Patent Application
20150178378
Current electronic user devices like smart phones and computers carry a plurality of functionalities, for example various programs for different needs and different modules for positioning, communication and entertainment. The variety of tasks that can be performed with these devices is large, and the individual tasks may be complex in nature. For example, a lot of the work carried out in today's world is done with the help of computers. Likewise, electronic devices have become part of everyday life in free time, as well. In addition, a large number of web services and applications are available through the internet for various purposes.
The user of a modern electronic device like a computer or a smart phone may be overwhelmed by the variety of functions provided by the system and by the variety of services and applications at his/her reach. For this purpose, the devices may provide ways of customizing the looks and the arrangement of functions in the device so that the functions of the device that the user needs are easily reachable. However, the state of the system as well as the situation where it is used may have an effect on the preference and needs of the user.
There is, therefore, a need for solutions for providing easier access to services and applications of the system.
Now there has been invented an improved method and technical equipment implementing the method, by which the above problems are alleviated. Various aspects of the invention include a method, an apparatus, a server, a client, a data structure and a computer readable medium comprising a computer program stored therein, which are characterized by what is stated in the independent claims. Various embodiments of the invention are disclosed in the dependent claims.
The example embodiments relate to forming a prediction using an experience matrix, a matrix based on sparse vectors such as random index vectors. At least a part of a first experience matrix and at least a part of at least a second experience matrix are caused to be combined to obtain a combined experience matrix. The experience matrices comprise sparse vectors or essentially similar vectors in nature, and said experience matrices comprise information of at least one system, for example contexts of a system. At least a part of at least one sparse vector of the combined experience matrix is accessed to form a prediction output, and a system is controlled in response to said prediction output.
In the following, various embodiments of the invention will be described in more detail with reference to the appended drawings of example embodiments, in which
a, 1b and 1c show representing words as basic sparse vectors, bags of words, and an experience matrix after updating with the bags of words,
a, 3b and 3c show providing a prediction by using an experience matrix, determining a difference of vectors in a context space, and providing a prediction by using an experience matrix,
a, 4b and 4c show representing bags of words as basic sparse vectors, updating an experience matrix by using a bag of words and basic sparse vectors, and an experience matrix after updating,
a and 5b show collecting words to a bag, and a situation where the words representing different instants of time,
a shows a system configured to gather and process data by using an experience matrix,
b shows a sensor arranged to provide words as the output,
c shows a sparse vector supply comprising a word hash table and a group of basic sparse vectors,
d shows a sparse vector supply comprising a group of basic sparse vectors,
e shows a sparse vector supply comprising a random number generator configured to generate basic sparse vectors,
f shows a system, which is controlled based on a prediction made by using an experience matrix,
a, 8b, 8c and 8d show bags of words, a user interface controlled based on a prediction, and method steps for preparing the system for a calendar event.
a, 9b and 9c show representing an experience matrix as a group of hash tables, updating the experience matrix, and using the experience matrix for making a prediction,
a and 10b shows generating auxiliary words by using a calendar and/or a dictionary, and replacing a word with a second word by using a dictionary,
a and 12b show different data processing devices,
a and 13b show various ways of combining two experience matrices or parts of experience matrices, and various resulting experience matrices (data structures),
In the following, several example embodiments of the invention will be described with reference to contexts of a system as well as a framework of random indexing. It is to be noted, however, that the invention is not limited to these uses or alternatives. In fact, the different embodiments have applications in any environment where semantic processsing and adaptation of the system is required.
Co-occurrence data may be gathered into an experience matrix based on co-occurrences of words in previous situations. The experience matrix may contain historical data about co-occurrences of words in different situations. The experience matrix EX1 may eventually contain a plurality of sparse vectors, which in turn contain information about historical co-occurrences of words related to a system. The matrix EX1 may be interpreted to memorize the experience gathered by one or more systems. The experience matrix may contain the “experience” gathered by one or more systems. The experience matrix may be used e.g. to “predict” suitable actions or operating modes which might be applicable in given situation when one or more (other) words related the situation are known. Thus, the system may change its operation based on observed facts, by using previous experience stored in the experience matrix of the system.
One or more prediction words may be determined from query word or words, by using the data stored in the experience matrix. The system may be subsequently controlled based on the prediction words. The query words may be words, which describe the current situation. The prediction words are words, which are likely to describe the current situation. Co-occurrence data may be stored as vectors of an experience matrix EX1. The co-occurrence data may be subsequently used for making a prediction. A vocabulary VOC1 may be needed to store co-occurrence data in the experience matrix EX1 and to utilize co-occurrence data stored in the experience matrix EX1. In an embodiment, the vocabulary VOC1 may be used also in connection with another matrix, called the basic matrix RM1.
Referring to
The vocabulary VOC1 may be a hash table, which indicates the location (e.g. row k) of the sparse vector in the basic sparse matrix RM1, by using the word (e.g. wk) as the key.
Each basic sparse vector of the matrix RM1 may represent a word. For example, the basic vector ak may be interpreted to represent the word wk in sparse vector format. Each basic sparse vector ak consists of elements R1,k, R2,k, . . . , Rik, . . . , Rm,k. In an embodiment, the basic sparse vectors a1, a2, . . . ak, . . . an of the matrix RM1 may be unique and different. Each row of the basic matrix RM1 may be a unique basic sparse vector associated with a different word. Each basic sparse vector ak may have a high number of zero elements and only a few non-zero elements. For example, a basic sparse vector ak may have e.g. 10000 elements R wherein twenty elements may be non-zero and 9980 elements may be zero.
In an embodiment, the sum of all elements of the basic sparse vector ak may be equal to zero. This may minimize memory consumption, may simplify mathematical operations and/or may increase data processing speed. In particular, 50% of the non-zero elements may be equal to −1 (minus one), and 50% of the non-zero elements may be equal to 1 (one). In other words, the value of an element may be −1, 0 or 1, and the basic sparse vector ak may be a ternary vector.
Each vector a1, a2, . . . ak, . . . an may represented by a point in a multi-dimensional space. More precisely, each vector a1, a2, . . . ak, . . . an may represented by a different end point in the same multi-dimensional space when the starting point of each vector is located at the same point (e.g. origin). The number m of elements R of the basic sparse vector ak may be e.g. in the range of 100 to 106. The number mnz of non-zero elements R of the basic sparse vector ak may be in the range of 0.1% to 10% of the number m, said number mnz of non-zero elements also being in the range of 4 to 103. Increasing the total number m and/or the number mnz of non-zero elements may allow using a larger vocabulary VOC1 and/or may provide more reliable predictions. However, increasing the number m and/or mnz may also require more memory space and more data processing power.
Rik denotes an element of the basic matrix RM1 belonging to the column and to the kth row. In an embodiment, the vectors may be ternary and the value of an individual element eik may be one of the following −1, 0, or 1. The number of negative non-zero elements R may be equal to the number of positive non-zero elements R, the values of the non-zero elements R being integers. This likely to maximize data processing speed, and to minimize the use of memory. However, this is not necessary. For example, a basic sparse vector ak may have sixteen elements R having a value −0.5 and four elements R having a value 2.
When the vectors a1, a2, . . . ak, . . . an of the matrix RM1 are compared with each other, the positions of the non-zero elements in a basic sparse vector ak and the values of the non-zero elements of the basic sparse vector ak may be randomly distributed. The basic sparse vectors may also be called as random index vectors. This randomness may ensure with a high probability that the points representing the basic sparse vector in the multi-dimensional space are not too close to each other. If two points representing two different words would be too close to each other, this might lead to erroneous predictions during subsequent processing. When the positions and the values are randomly distributed, this may also ensure with a high probability that each word of the vocabulary VOC1 is associated with a unique and different basic sparse vector. When the indices are random, it is highly probable that the elementary sparse vectors associated with two different words are orthogonal or nearly orthogonal. Thus, the dot product of said elementary sparse vectors is equal to zero at a high probability. This pseudo-orthogonality of the words may preserve the unique identity of each word or event or unique occurrence stored in the matrix, even when they are represented by the sparse vectors. This pseudo-orthogonality of the words may preserve the unique identity of each word, event or unique occurrence stored in the experience matrix EX1 when the words, events and/or occurrences are represented as combinations of the sparse vectors in the experience matrix EX1. The words of the vocabulary may be arranged e.g. in alphabetical order.
b shows several sets of words, which may be called e.g. as “bags”. A bag comprises two or more different words related to each other. The number of words in a bag may be e.g. in the range of 2 to 1000. Each bag could also be called as a “document”.
The bags represent co-occurrences of words. The bags may be used to educate a system. Information about co-occurrences of the words may be stored in a matrix EX1 (
The words of a bag may e.g. describe a situation and/or a system used in said situation. In particular, the words may represent observed facts.
For example, a first bag BAG1 may comprise the words “work”, “morning” and “mail”. The word “work” may be e.g. determined e.g. based on a marking in the calendar of the system. The word “morning” may be provided e.g. by a clock of the system. The word “mail” may be related to the words “work” and “morning” e.g. because the system has detected that a user has received and/or sent written messages in a situation described by the words “work” and “morning”.
A second bag BAG2 may comprise the words “home”, “meal” and “music”. The word “home” may be provided e.g. when a navigation unit of carried by a user indicates that the detected location matches with a location called as “home”. The word “meal” may be provided e.g. because food (e.g. pizza) has just been ordered to the location “home” via a delivery service. The system may be arranged to provide a word “music” e.g. because user has instructed the system to play music.
The words home, meal and music may represent e.g. a situation prevailing at a first instant t1. The words home, meal and music might be inter-related. If this combination of words has occurred several times in the past, it may e.g. used to evaluate (predict) the likelihood at which the user wishes to listen to music when eating meal at home. For example, a portable device may be arranged to set itself into a music playing mode when it determines that the user is eating meal at home.
A third bag BAG3 may comprise the words “work”, “meeting” and “silent”. The words “work” and “meeting” may be e.g. determined e.g. based on markings in the calendar of the system. The system may be arranged to provide a status descriptor “work” e.g. based on information contained in a calendar of the user. The system may determine the word “work” also when the location of a portable device of the user matches with the location of the working place. The word “silent” may be derived e.g. because a user attending a meeting has set the system to a silent mode. In particular, the user may have set a mobile phone to an operating mode where the phone does not generate an audible alarm in case of an incoming telephone call.
Now, the words work, meeting and silent may represent a situation prevailing at a second time t2. If this combination of words has occurred several times in the past, it may e.g. used to evaluate (predict) the likelihood at which the user wishes to set his portable to the silent mode when attending a meeting at work. For example, the portable device of the user may now be arranged to automatically suggest setting into a silent mode when it determines that the user is attending a meeting.
Referring to
The vocabulary VOC1 may be the same (hash table) as in
Each vector of the experience matrix EX1 may be a linear combination of sparse vectors and an initial vector.
In particular, each vector of the experience matrix EX1 may be a linear combination of the basic sparse vectors of the basic sparse matrix RM1 (
The number n of elements eik of an individual vector vk may be equal to the number n of elements R of an individual basic sparse vector shown in
The experience matrix EX1 may contain co-occurrence data obtained from a plurality of bags BAG1, BAG2. When the co-occurrence data obtained from the bags is stored in the experience matrix EX1 as linear combinations of the sparse vectors, this may represent storing the co-occurrence data in a compressed format.
In an embodiment, the number n of elements eik of an individual vector vk may be substantially smaller than the number of bags from which the co-occurrence data stored in the matrix EX1 was obtained.
It is not necessary to change the number n of elements eik when new data is added from a new bag. This means that the matrix EX1 may be updated fast and easily by using co-occurrence data obtained from further bags. This also means that the size (i.e. the number of rows and columns) of the matrix EX1 may remain constant even when the matrix EX1 is updated by using co-occurrence data obtained from the further bags.
However, in an embodiment, it might be impossible to unambiguously determine (by using only the vocabulary VOC1 and the matrices RM1, EX1) which ones of the bags used for updating contained a given word. In this sense, storing co-occurrence data of the bags in the matrix EX1 may represent a lossy way of storing co-occurrence data.
The experience matrix EX1 may be initially a zero matrix (null matrix) but this is not necessary. The initial values of several vectors of the experience matrix EX1 may exhibit smooth spatial variation.
The symbols W1, W2, W3 (with uppercase letter) denote the words of the bag BAG1. The w1, w2, . . . wn (with lowercase letter) denote the words of the vocabulary VOC1. Typically, the first word W1 of the bag is not the same as the first word w1 of the vocabulary, but sometimes the word W1 might perhaps be the same as the word w1.
A vector of the experience matrix EX1 may be updated based on words contained in a bag. During the updating, the vector of the experience matrix EX1 may be called e.g. as a target vector.
The updating may comprise modifying the target vector (e.g. vP) associated with a word (e.g. “mail”) of the bag BAG1. A bag contains two or more words W1, W2, W3, . . . When the words W1, W2, W3 belong to the same bag BAG1, this indicates, among other things, that the words W1 and W2 belong to the context of the word W3.
One or more other words belonging to a bag together with a first word may be considered to belong to the context of the first word. A word may represent e.g. a state of a system and/or an external condition.
The target vector vP of the experience matrix EX1 associated with the word W3 of the bag BAG1 may be identified by using the vocabulary VOC1. More precisely, the location of the vector vP (row) of the experience matrix EX1 associated with the value (=wp, “mail”) of the word W3 may be determined by using the vocabulary VOC1.
The basic sparse vectors (e.g. ak+1 and an−1) representing the other words W1, W2 of the bag BAG1 may also be found by using the vocabulary VOC1.
Adding co-occurrence data of a bag BAG1 to the experience matrix EX1 may comprise replacing the (first) target vector vP associated with the word W3 of the bag BAG1 with a modified vector which is a linear combination of the (first) target vector vP and the basic sparse vectors ak+1, an−1 representing the other words W1, W2 of the bag BAG1.
In an embodiment, the basic sparse vectors ak+1, an−1 may be simply summed to the target vector vP. In other words, the weighting coefficient of each basic sparse vectors ak+1, an−1 used in the linear combination may be equal to one. However, the weighting coefficient of a basic sparse vector may also deviate from one in order to emphasize or reduce the contribution of the word represented by said basic vector. The weighting coefficient may be e.g. in the range of 0.1 to 10.
A second target vector (e.g. vn−1 shown in
If the bag contains more than three words, a target vector associated with each additional word may be updated by adding contribution of the other words of the bag, respectively.
The matrix may be updated by using words of a second bag, and by using words of any subsequent bags, respectively.
The two or more words W1, W2, W3, . . . of a bag BAG1 may be provided by a system (system words) and/or by a client (client words). The system words, may be e.g. words describing time and/or location. The “client” may be e.g. a human user, an application or a network service. The client words may be e.g. words obtained from a human user and/or from a running software application. A word may e.g. describe the state of a system.
a shows making a prediction by using the experience matrix EX1. The experience matrix EX1 may be e.g. the matrix shown in
One or more words representing a state of the system may be predicted by using the experience matrix EX1. The prediction method may comprise receiving one or more query words Q1, Q2, . . . . (
The query vector QV1 may now be compared with the vectors v1, v2, . . . vn of the experience matrix EX1 in order to identify one or more vectors v1, v2, . . . vn of the experience matrix EX1 which are substantially similar to the query vector QV1. During the comparing, the vectors v1, v2, . . . vn of the experience matrix EX1 may be called as candidate vectors. The word or words associated with vectors having the smallest difference to the query vector QV1 may be determined and provided as the prediction output OUT1. The prediction method may comprise determining a difference between the query vector QV1 and each vector v1, v2, . . . vn of the experience matrix EX1.
b illustrates the query vector QV1 and a few vectors of the experience matrix EX1 in the multidimensional space when the starting points of the vectors coincide (e.g. at the origin ORIG1). DIM and DIM′ denote two dimensions of the multidimensional space (as discussed above, the number of dimensions of the space may be e.g. 10 000 or higher). If two vectors have at least approximately the same length and at least approximately the same direction in the multidimensional space, this may be interpreted to indicate that said two vectors are associated with two words appearing in substantially similar context. If one of said words is a valid descriptor for a state of the system, it is likely that also the other word is a valid descriptor for a state of the system.
DV(QV1,v) denotes a difference vector between the query vector QV1 and a vector v of the experience matrix EX1. A value indicative of a difference between the query vector QV1 and a vector v of the experience matrix EX1 may be determined based on the query vector QV1 and a vector v of the experience matrix EX1. The “difference” may expressed e.g. by an angle θ between the query vector QV1 and the vector v of the experience matrix EX1. The “difference” may expressed e.g. by the cosine cos(θ) of the angle θ between the query vector QV1 and the vector v of the experience matrix EX1. The cosine (cos(θ)) of the angle θ between two vectors v and QV1 may be calculated from the dot product, by using the equation |v∥QV1| cos(θ)=v·QV1. The difference may expressed e.g. by the Euclidean length |DV| of the difference vector DV. The difference may expressed e.g. by the dot product QV1·v of the vectors. The difference may expressed e.g. by the square |DV|2 of the Euclidean length |DV| of the difference vector DV. The difference may expressed e.g. as the taxicab length (Manhattan length) of the difference vector DV. Also other metrics may be used when defining the length of the difference vector DV. Also other metrics may be used when evaluating the difference between the vectors.
One or more prediction output words may now be selected from the vocabulary VOC1 based on the differences between the query vector QV1 and the vectors of the experience matrix EX1. If the difference between the query vector QV1 and a candidate vector fulfils a selection criterion, the word associated with said candidate vector may be provided as a prediction word.
The differences DV may be compared with each other in order to find the sparse vector or vectors which deviate by the smallest distance DV from the query vector QV1. In particular, the differences DV may be sorted. A word associated with a sparse vector having the smallest distance DV may be provided as the output OUT1 of the prediction method. The output may be called as the prediction. Also several words associated with sparse vectors having the smallest distance DV may be provided as the output OUT1. For example, NP words associated with candidate vectors having the smallest distance DV may be provided as the output OUT1, wherein the integer NP may be e.g. in the range of 2 to 10. In the example of
The query words Q1, Q2 are already known to describe the situation. The query words Q1, Q2 may be excluded from the prediction words OUT1. Vectors associated with the query words Q1, Q2 may be excluded from the comparison when the query vector QV1 is compared with the other vectors of the experience matrix EX1.
The prediction word or words may be interpreted to be an answer to a conditional question “which prediction word or words OUT1 are likely to describe a situation if the situation is already described by the one or more (known) query words Q1, Q2. In the example of
One or more of these prediction words may be used for controlling a system. For example, a system could automatically activate an application for receiving and sending e-mails in a situation, which can be described by the word “morning”. For example, a system could automatically set itself into a silent mode of operation in a situation, which can be described by the word “morning”.
The one or more prediction words could be communicated back to a client making the query. The one or more prediction words could be communicated to a different client. The system may be arranged to utilize all prediction words for the control, or the system may be arranged to utilize a smaller subset of the prediction words for the control, wherein one or more prediction words may be ignored.
c shows making a query by using two or more query words Q1, Q2. Increasing the number of query words Q1, Q2 of a single query may improve the probability that the prediction word or words will be valid descriptors for the situation. In this example, the query consists of two words Q1, Q2. A first vector vk of the experience matrix EX1 associated with the query word Q1 (e.g. morning) may be found by using the vocabulary VOC1. A second vector vn−1 associated with the query word Q2 (e.g. work) may be found by using the vocabulary VOC1. The query vector QV1 may now be determined as a linear combination of the vectors vk, vn−1 associated with the query words Q1, Q2. In particular, the query vector QV1 may be determined as the sum of the vectors vk, vn−1. Alternatively, the query vector QV1 may be formed e.g. by selecting the elements from the vectors Q1 and Q2 having maximum absolute value into the query vector QV1, or by another method of combining. The prediction words may be subsequently determined by comparing the query vector QV1 with the vectors v1, v2, . . . vn of the experience matrix EX1, as described above for the case of
The prediction words may be subsequently formed by a method comprising:
Vectors vk+1, vn−1 of the experience matrix EX1 associated with the query words Q1, Q2 may be excluded from the comparison.
In this case, the vector vP of the experience matrix EX1 appears to provide a difference vector DV having the shortest length (|DV|2=2). In this case, the prediction word corresponding to the vector vP is “mail”. Also several words associated with NP vectors having the shortest distance vector DV to the query vector may be provided as the prediction. The number NP may be e.g. three. In this case, the words “mail”, “meeting” and “silent” may be provided as the prediction words.
The updating method shown in
a shows three bags BAG1, BAG2, BAG3 of words. The bags may be e.g. the same as in
b shows updating the experience matrix EX1 by using co-occurrence data derived from the bag BAG1. The bag contains two or more words W1, W2, W3 (e.g. the words “morning”, work” and “mail”). Each word W1, W2, W3 of the bag BAG1 is now represented by the same basic sparse vector (e.g. b1). The words W1, W2, W3 of the bag BAG1 are the same as the words wP, wk+1, wn−1 of the vocabulary VOC1. This allows finding the vectors vP, vk+1, vn−1 of the experience matrix EX1 associated with the words of the bag BAG1. These vectors may be called as the target vectors.
The experience matrix EX1 may be initially a zero matrix (null matrix) but this is not necessary. The initial values of several vectors of the experience matrix EX1 may exhibit smooth spatial variation. The matrix EX may already contain vectors obtained as linear combinations of previous vectors of the matrix EX1 and sparse vectors. The experience matrix may in the latter situation be such that it is not saturated, that is, new data that is added to the matrix can still be reliably retrieved. An experience matrix having the property of not being saturated may be understood to be sparse in the sense of the present embodiments even though a high number of its elements deviate from zero.
The basic sparse vector b1 representing the bag BAG1 may now be added to each target vector vP, vk+1, vn−1 associated with a word of the bag BAG1. The same basic sparse vector b1 is now common to all words of the bag BAG1. The same basic sparse vector b1 may be added to all target vectors vP, vk+1, vn−1 associated with the words of the bag BAG1. In general, a modified target vector associated with a word of the bag BAG1 may be formed as a linear combination of the target vector associated with said word and the basic sparse vector b1. The second bag BAG2 may be represented by a different basic sparse vector (e.g. b2). The basic sparse vector b2 representing the second bag BAG2 may now be added to each target vector associated with a word of the bag BAG2. The third bag BAG3 may be represented by a different basic sparse vector (e.g. b3). The basic sparse vector b3 representing the third bag BAG3 may now be added to each target vector associated with a word of the bag BAG3.
Thus, the experience matrix EX1 may be updated by using contents of a plurality of bags BAG1, BAG2, BAG3. Several vectors of the experience matrix EX1 may eventually be linear combinations of different sparse vectors b1, b2, b3.
c shows an experience matrix EX1 obtained after the co-occurrence data of the bags BAG1, BAG2, BAG3 has been stored in the experience matrix EX1 by using different basic sparse vectors b3 representing the bags. The sum of all elements eik of the matrix EX1 may be equal to zero. eik denotes an element of the experience matrix EX1 belonging to the column and to the kth row. The value of the element eik may be an integer in order to maximize data processing speed. The value of the element eik is not limited to the range −1 to 1.
The experience matrix EX1 determined by using the method shown in
a-5c show collecting words from a buffer to a single bag. Words may be received from one or more clients and/or from the system. The received words may be (temporarily) stored in one or more buffers. The words may be collected from the buffers to a bag. The experience matrix may be updated by using the contents of the bag. The experience matrix may be subsequently updated by using the contents of a second (different) bag.
In
In
In
The method of
The word W1,t1 could describe e.g. the weather conditions at the time t1 (e.g. “raining” at the time 14:00), and the word W2,t2 could describe e.g. a marking in the calendar of a user (e.g. “sports” at 18:00). The prevailing weather conditions may be received e.g. from an internet service. In this case the weather conditions prevailing at 14:00 might have a causal effect on whether the user will be truly practicing sports outside at 18:00. In this sense, it might be useful to collect the earlier word “raining” and the later word “sports” into the same bag.
The term “pull” may mean that communication of data from the sensor to the buffer may take place such that the communication is not initiated by the sensor. In an embodiment, the sensor may provide a system word whenever requested. In an embodiment, communication of data from the sensor does not start updating the experience matrix.
The term “push” may mean that communication of data from a client to the buffer may be initiated by the client. In an embodiment, communication of data from the client may also initiate updating the experience matrix.
The term “flush” may mean updating the experience matrix by using one or more bags of words. The bags may be collected from the contents of the buffer.
The experience matrix may also be used when providing a prediction, as described above (See the discussion related to
a shows a subsystem 400 for processing co-occurrence data. The subsystem 400 may be arranged to store co-occurrence data in an experience matrix EX1. The subsystem 400 may be arranged to provide a prediction based on co-occurrence data stored in the experience matrix EX1.
The subsystem 400 may comprise a buffer BUF1 for receiving and storing words, a collecting unit WRU for collecting words to a bag, a memory MEM1 for storing words of the bag, a sparse vector supply SUP1 for providing basic sparse vectors, memory MEM3 for storing the vocabulary VOC1, the vocabulary stored in the memory MEM3, a combining unit LCU1 for modifying vectors of the experience matrix EX1 and/or for forming a query vector QV1, a memory MEM2 for storing the experience matrix EX1, the experience matrix EX1 stored in the memory MEM2, a memory MEM4 for storing the query vector QV1, and/or a difference analysis unit DAU1 for comparing the query vector QV1 with the vectors of the experience matrix EX1. The subsystem 400 may optionally comprise a buffer BUF2 and or a buffer BUF3 for storing a query Q1 and/or a prediction OUT1.
The words may be received from one or more sensors NAV1, G1, from a clock CLK1 and/or from one or more clients C1. The words may be collected to individual bags by a collector unit WRU1. The words of a bag may be collected (i.e. temporarily stored) in the memory MEM1. The contents of each bag may be communicated from the memory MEM1 to a sparse vector supply SUP1. The sparse vector supply SUP1 may be configured to provide basic sparse vectors for updating the experience matrix EX1.
The contents of each bag and the basic sparse vectors may be communicated to a combining unit LCU1, which may be configured to modify the vectors of the experience matrix EX1 (e.g. by forming a linear combination). In particular, the combining unit LCU1 may be configured to add basic sparse vectors to target vectors specified by the words of each bag. The combination unit LCU1 may be arranged to execute summing of vectors at the hardware level. Electrical and/or optical circuitry of the combination unit LCU1 may be arranged to simultaneously modify several target vectors associated with words of a single bag. This may allow high data processing rate.
The experience matrix EX1 may be stored in the memory MEM2. The words may be associated with the vectors of the experience matrix EX1 by using the vocabulary VOC1 stored in the memory MEM3. Also the vector supply SUP1 may be configured to use the vocabulary VOC1 (or a different vocabulary) e.g. in order to provide basic sparse vectors associated with words of a bag.
The subsystem 400 may comprise the combining unit LCU1 or a further combining unit configured to form a query vector QV1 based words of a query Q1. They query vector QV1 may be formed as a linear combination of vectors of the experience matrix EX1. The locations of the relevant vectors of the experience matrix EX1 may be found by using the vocabulary VOC1. The query vector QV1 may be stored in the memory MEM4.
The difference analysis unit DAU1 may be configured to compare the query vector QV1 with vectors of the experience matrix EX1. A difference analysis unit DAU1 may be arranged to determine a difference between a vector of the experience matrix EX1 and the query vector QV1. The difference analysis unit DAU1 may be arranged to sort differences determined for several vectors. The difference analysis unit DAU1 may be configured to provide a prediction OUT1 based on said comparison. The query words Q1, Q2 itself may be excluded from the prediction words.
The difference analysis unit DAU1 may be arranged to compare the vectors at the hardware level. Electrical and/or optical circuitry of the combination unit LCU1 may be arranged to simultaneously determine quantitative difference descriptors (DV) for several vectors of the experience matrix EX1. This may allow high data processing rate.
The subsystem 400 may comprise a control unit CNT for controlling operation of the subsystem 400 and/or for controlling operation of the system 500 (see
Words may be received e.g. from one or more clients C1, from one or more sensors G1, from a position sensor NAV1, and/or from a clock CLK1. The prediction word or words OUT1 may be communicated to the client C1 or to a different client. The client C1 may also retrieve system words from the buffer BUF1 e.g. in order to form a query Q1.
b shows, by way of example, how a word may be provided based on physical interaction between a sensor and its environment. A sensor G1 may provide one or more electrical or optical sensor signals S0 by monitoring its physical environment. The monitoring comprises passive or active interaction with the physical environment. For example, in case of passive interaction, a microphone or a light sensor may receive sound or light waves (which are not initiated by the system). The signal S0 may be e.g. an analog signal representing the magnitude and frequency of a sound wave received by a microphone. The signal S0 may be e.g. an analog signal representing the magnitude and frequency of vibrations detected by an acceleration sensor.
In case of active interaction, the sensor G1 may transmit e.g. ultrasound or infrared light, which may be reflected from a surface of the environment. For example, the distance to the surface may be estimated e.g. based on the reflected sound or light waves received by a microphone or a light sensor. Optionally, a digital sensor signal S1 may be provided by processing the (analog) signal S0 in a sensor interface GIN1. The digital sensor signal S1 may comprise e.g. spectral analysis of a sound received by a microphone. The digital sensor signal S1 may comprise a digital image captured by an image sensor.
Some sensors may directly provide a quantized digital signal, which may be directly used as a system word. For example, a microswitch proximity sensor may or a digital encoder of a joystick may directly provide digital words representing state of a system. For example, a microswitch may provide either a word “1” or a word “0” specifying whether something is pressing the switch or not.
Optionally, a quantizer QU1 may be arranged to provide a quantized sensor signal S2 based on the digital sensor signal S1. The quantization may be performed by using a mapping function, which may be modifier later, if needed. For example, sound volume values ranging from 0 to 127 (128 different values) may be mapped into a smaller number of values (e.g. 0, 1, 2).
Optionally, the quantizer QU1 may comprise an analyzer, which is arranged to classify the signal S1 into two or more different classes. The quantizer QU1 may be a classifier. The quantizer QU1 may provide a value specifying the class of the signal S1. For example, the quantizer QU1 may be arranged to provide values, which specify the class of a sound based on spectral analysis (e.g. “0”=motor, “1” music, “2” human voice).
A value provided by the quantizer QU1 may be used as an input word for the subsystem 400. Optionally the value may be converted to a (different) word e.g. in order to make it more easily intelligible to a human user.
The quantized sensor signal S2 may be optionally converted into a different word which describes the physical environment of the sensor in a natural language (e.g. in the English language or in Finnish language). The signal S2 may be converted into the word by using a dictionary. The contents of the dictionary may be fixed or it may be updated at a later stage. Numerical words (0,1,2) describing the nature of a sound may be converted into words of a natural language e.g. by using the following dictionary “0”-“motor”, “1”-“music”, “2”-“voice”. Numerical words (0,1,2) describing the volume of a sound may be converted into words of a natural language e.g. by using the following dictionary “0”-“silent”, “1”-“normal”, “2”-“loud”.
Optionally, a marker unit MRK1 may be arranged to modify a word by adding an identifier, which makes the word unique and different from other words. The identifier may e.g. specify the identity of the sensor S1 or the type of the quantity specified by said word. For example, when the word is a numerical word (e.g. “0”, “1” and “2”) describing sound volume, a prefix “volume” may be attached to the word so as to form e.g. a word “volume 0”.
Thus, a sensor S1 may be arranged to provide a sensor word (system word) by using one or more method steps described above.
In an embodiment, an analysis unit receiving a signal from an acceleration sensor may be arranged to determine e.g. whether a portable device is stationary, carried by a walking person, carried by a running person, or carried in a moving vehicle. The analysis unit may provide words, e.g. “stationary”, “walking”, “running”, “traveling” describing these different states. The acceleration sensor may be e.g. a capacitive or a piezoelectric acceleration sensor.
However, the signal obtained from sensor G1 does not need to be converted into a sensor word having a clear semantic meaning, i.e. it is not necessary to use the dictionary THESA1. In an embodiment, an analysis unit receiving a signal from an acceleration sensor may be arranged to provide sensor words describing different acceleration conditions, e.g. “acceleration A”, “acceleration B”, and “acceleration C”. A system 500 comprising the subsystem 400 and the sensor G1 may be operated such that the system 500 can make successful predictions even if the semantic meaning of a sensor word is not known.
A word provided by a sensor G1 may be presented e.g. as a uniform resource identifier (URI) containing a scheme and a scheme specific part. For example, a word provided by using an acceleration sensor may be e.g. “motion://moving” or “motion://stationary”, where the first part “motion” is the scheme, and the second part “moving” or “stationary” is the scheme specific part. The parts may be separated by the symbols “://”. A word provided by using the acceleration sensor may also be e.g. “orientation://vertical” or “orientation://horizontal”.
Referring to
Referring to
Alternatively, or in addition, an individual basic sparse vector associated with a word may be generated in real time when said word is used for the first time in a bag. The basic sparse vectors may be generated e.g. by a random number generator. The configuration of
Referring to
Referring to
Referring to
The system 500 may optionally comprise one or more sensors G1 for providing system words describing e.g. status of the system or a property of an environment of the system. The system 500 may optionally comprise a clock CLK1 for providing time data. The system 500 may optionally comprise a location detector NAV1 for providing location data. The location detector may be e.g. a GPS navigation unit.
Co-occurrence data may be stored in the experience matrix EX1 as described above.
A client C1 may send a query Q1 to the subsystem 400. The query Q1 may comprise e.g. a word provided by a human user via a user interface 501, from an application APP1, and/or from a sensor G1 of the system 500. The query Q1 may comprise e.g. a word retrieved from an internet service (e.g. weather information). The subsystem 400 may provide one or more prediction words OUT1 based on the query Q1, as described above. The system 500 may be configured to start and/or stop an application APP1, APP2 based on a prediction OUT1. The operation of the system 500 may be controlled based on the prediction OUT1. The prediction OUT1 may be used e.g. as an input for a web browser. The system 500 may be configured to display information on a display of a user interface 501 based on the prediction OUT1 (user interface management). The system 500 may be configured to set an operating mode based on the prediction OUT1. The system 500 may be configured to allocated data processing power to different applications and/or units based on the prediction OUT1. The system 500 may be configured to start and/or stop operation of a functional (physical) unit of the system (e.g. a GPS navigation unit or an audio amplifier) based on the prediction OUT1.
Operation of a first device of the system 500 may be controlled based on words obtained from a second device of the system 500. For example, the first device may be carried by the user (a first person), and the second device may be carried by a second person. The second person may be e.g. a family member, a friend, or a workmate of the first person. The query Q1 may comprise one or more words related to the status of the second device and/or related to the calendar of the second person. The system may now provide a prediction OUT1 based on the query Q1. The system may propose activation of a functionality based on the prediction OUT1.
For example, if the second device has been set into the silent mode, the system, 500 may suggest based on the prediction that the preferred way of contacting the second person might be e.g. a SMS message or e-mail message, instead of making a phone call. The experience matrix EX1 may store co-occurrence data, which indicates that an attempt to make a phone call has almost never been successful when the second device has been set into the silent mode. The experience matrix EX1 may store co-occurrence data, which indicates that the user of the second device has sometimes immediately responded to an e-mail message even when the second device has been set into the silent mode. Thus, when the user of the first device attempts to make a phone call to the second person, the system 500 may propose activating an application for sending messages to the second device (e.g. SMS message or e-mail).
a-8c show an example for controlling operation of a communications system by using predictions.
The experience matrix EX of the system 500 may be updated by using the bags BAG4 to BAG9 shown in
Referring to
When the meeting is about to start (e.g. at the time 8:59), the system 500 may propose changing the operation mode (“profile”) to the silent mode e.g. in order to avoid disturbing the meeting with incoming telephone calls. One or more information fields F1-F4 may be displayed to the user. A field F1 may propose changing to the silent mode. The user may accept or reject the proposition e.g. by touching the (virtual) key(s) B1 or B2. In an embodiment the key B2 may be omitted. If the user does not touch the key B2 within a predetermined time period, the system may automatically change to the silent mode. In an embodiment the key B1 may be omitted. If the user does not touch the key B1, the system may continue operation in the normal mode. In an embodiment both keys B1, B2 may be omitted. The user may have previously authorized the system to set the operating mode automatically based on the prediction.
The fields F2-F4 may display auxiliary information. The field F3 may optionally display e.g. the actual time. The field F2 may optionally display e.g. the next relevant event (e.g. meeting with person A). The field F4 may optionally display e.g. a calendar view. In addition to setting the operating mode, the system 500 may also control displaying fields F1-F4, and/or displaying the keys B1, B2 based on the prediction.
Thanks to controlling operation based on the prediction, the user may now set the system to the silent mode by touching one key only. In an embodiment, the user does not need to do anything in order to set the system to the silent mode.
Without the prediction, the user might need to seek a suitable menu from the menu structure of the system in order to set the system to the silent mode. Typically, this operation would require making and confirming several selections in the menus of the menu structure of the user interface 501.
Referring to
Based on the experience, the words “meeting” and “Person B” often appear together with the word “teleconference with person C”. This co-occurrence data has already been stored in the experience matrix EX1.
A client of the system 500 may form a query Q1 containing the words “9:00”, “meeting”, “Person B”. The subsystem 400 may provide a prediction based on the query Q1″ by using the experience matrix EX1 comprising the co-occurrence data of
If the user accepts the proposition, the system may automatically start a teleconference. This may comprise automatically activating a camera module for capturing video images of the user. This may comprise (automatically) opening a video communication link and/or audio communication link with an interface of the person C at a remote location.
In an embodiment the key B2 may be omitted. If the user does not touch the key B2 within a predetermined time period, the system may automatically open the telecommunication link. In an embodiment the key B1 may be omitted. If the user does not touch the key B1, the system does not open the communication link. In an embodiment both keys B1, B2 may be omitted. The user may have previously authorized the system to open the telecommunication link automatically.
Thanks to controlling operation based on the prediction, the user may now open the communication link by touching one key only. In an embodiment, the user does not need to do anything in order to open the communication link.
Without the prediction, the user might need to seek a suitable menu from the menu structure of the system in order to open the telecommunication link. Typically, this operation would require making and confirming several selections in the menus of the menu structure of the user interface 501.
d shows method steps for updating the experience matrix EX and for controlling the system based on the prediction. In step 705, a scheduled event or an actual event may be stored in a memory of the system. In particular a calendar event may be added to the memory e.g. via the interface 501 by a user or by another person. In step 710, the user or another person may set an operating mode of the system, associated with the event mentioned in step 705. In step 715, the user or another person may start and/or stop one or more applications, associated with the event mentioned in step 705. In step 720, words describing the event, operating mode and/or applications may be collected into a bag. An experience matrix EX1 of the system 500 may be updated by using the bag. Co-occurrence data representing previous events may already be stored as the (sparse) vectors of the matrix EX1 before this step, i.e. the matrix does not need to empty prior to step 720. The steps 705-720 may be repeated several times in order to gather co-occurrence data related to several different events.
In step 750, the system may form a query Q1 comprising a word describing an event. The query Q1 may also comprise a word specifying additional information associated with the event, e.g. an identifier of a person, who is associated with the event. In step 755, the subsystem 400 may provide a prediction by using the query Q1 and the vectors previously stored in experience matrix EX1. In step 760, the system 500 may e.g. propose starting and/or stopping one or more applications and/or it may propose setting an operating mode based on the prediction. In step 765, the system may request permission from the user to carry out the action proposed in step 760. In an embodiment, the user may have given the permission in advance. In step 770, the response given by the user may recorder in a memory and used for updating the experience matrix EX1. In particular, a word describing the response may be collected in the same bag together with one or more other words related to the query Q1. The words of the bag may be used for updating the experience matrix EX1. In step 775, the operating mode may be set according to the prediction, if this was accepted by the user in step 765. In step 780, one or more applications may be started and/or stopped according to the prediction, if this was accepted by the user in step 765.
The steps 705-780 may be repeated several times. The steps 705-780 may be repeated e.g. several times per day during the lifetime of a portable device belonging to the system 500. The lifetime may be several years, e.g. more than three years. Thus, the experience matrix EX1 may eventually contain co-occurrence data representing e.g. more than 1000 events (data representing more than 1000 bags).
The memory MEM2 containing the experience matrix EX1 may be included in a first portable device carried by a user. The first portable device may be called e.g. as a smart phone and/or as a portable computer (which may have wireless communication capabilities). The first portable device may be e.g. one of the devices 1260, 1251, 1250 shown in
Sooner or later, the first portable device may become outdated, it may be lost, or it may be broken. A (back-up) copy of the experience matrix EX1 may be stored in the system 500 in a location, which is remote and/or separate from the location of the memory MEM2. The experience matrix EX1 containing the previous data may be easily copied to one or more portable devices, if needed.
An experience matrix EX1 may be copied from the device of a first person to a device of a second person. An experience matrix EX1 may be transferred or copied from one device to another device e.g. via internet, via en electrical cable, via an optical cable, via a radio frequency link, via a mobile telephone network, or by using a physical memory (i.e. a computer memory product) storing the experience matrix EX1. The copying may be associated with a payment, i.e. the experience matrix EX1 may also be sold as an individual item, separately from the original system 500 which was used for collecting the data into the matrix EX1. An experience matrix EX1 may be wholly or partially transferred from one device to another, and may, but does not have to, replace partly or wholly the destination matrix. For example, a user may receive an update of the experience matrix related to a vacation trip, such as information regarding to Italian Alps.
The data stored in the experience matrix EX1 is likely to be useful when the queries Q1 contain same or like words which were used for updating the experience matrix EX1. Previous data stored in an experience matrix EX1 may be applicable only in a certain type of systems, wherein the data may be substantially useless in a second type of a system, which uses different query words and/or operates in a fundamentally different environment. For example, an experience matrix EX1 which only comprises data related to professional meetings is not likely to be useful for predicting which video clips the user wishes to view in his free time.
a shows how an experience matrix EX may be represented by using hash tables. The majority of the elements eik of an experience matrix EX1 may be zero elements. Several vectors of the experience matrix EX1 may be zero vectors. In that case, it may be needless to use memory space for storing the zero values. The vocabulary VOC1 may be a hash table, wherein each word stored in the vocabulary VOC1 may be used as the key. Entries in the vocabulary VOC1 point to value arrays INDH, which contain the actual values of the non-zero elements eik of the experience matrix EX1. For example, an individual word wk may point to a value array INDHk, which contains values of the non-zero elements of the vector vk of the experience matrix EX, said vector vk being associated with the word wk. All elements of the value array INDHk may be non-zero.
A second hash table CAH1 may be used to keep track of those columns of the experience matrix, which contain non-zero elements. The second hash table CAH1 may be called e.g. as a “columns activated hash table”. The second hash table CAH1 may have the column number (i) as the key. Entries in the second hash table point to occurrence arrays OCCL, which in turn may contain a list of row numbers (k) of words (wk) contributing to each column of the experience matrix EX1. For example, an occurrence array OCCLi may contain the row numbers k1, k2, k3, to indicate that the elements e at the column i and at the rows k1, k2, k3 are non-zero.
The hash tables VOC1, CAH1, the group of the value arrays INDH and the group of the occurrence arrays OCCL may together fully specify the values and the locations of the non-zero elements of the experience matrix EX1.
A value array INDH may contain only the values of the non-zero elements of a single vector of the experience matrix EX1. It does not need to specify the indices of the columns of the experience matrix EX1 containing said non-zero values, because this information may be derived from the occurrence arrays pointing to the word associated with said value array INDH. However, the value array INDH may contain additional information in addition to the values of the non-zero elements, e.g. the indices of the columns of the experience matrix EX1 containing the non-zero values. The value arrays INDH may together form a value matrix. The occurrence arrays OCCL may together form an occurrence matrix.
b shows method steps for updating an experience matrix EX1 represented by the hash tables. In this case, a predetermined group of words contained in the vocabulary VOC1 may be used in order to avoid changing the number of rows of the experience matrix EX1. Words not belonging to said predetermined vocabulary VOC1 may be ignored. In step 810, a bag of words may be received. In step 830, it may be checked whether a word of the bag is contained in the vocabulary VOC1. Words not contained in the vocabulary VOC1 may be ignored in step 820 in order to avoid changing the number of rows of the experience matrix EX1. In step 840, a basic sparse vector representing each word of a bag may be determined by using the vocabulary VOC1 and the basic sparse matrix RM1 (see
The number of non-zero elements of a vector of the experience matrix EX1 may also need to be increased during the updating. Thus, the length of a value array INDHk may be increased in order to store the additional values. The locations of the values stored in value array INDHk may be in the same order as in the vector of the experience matrix EX1 represented by said value array. This means that an element may be added also in the middle of a value array INDHk and/or to an end of the value array INDHk.
In step 860, the hash table CAH1 and the occurrence arrays OCCL may be updated according to the new elements, which were possibly added to the value array INDH. If a non-zero value was stored in an element eik in a column (i), which was previously a zero column, the hash table CAH1 may be updated to contain a pointer to an occurrence array OCCL, which in turn contains one or more pointers to the words where the column (i) was used. The hash tables VOC1, CAH1 and the arrays INDH and OCCL contain the co-occurrence data of an experience matrix EX1. An experience matrix EX1 may be represented in a hash table format, by using the hash tables VOC1, CAH1 and the arrays INDH and OCCL.
c shows method steps for making a prediction by using the experience matrix EX1. The experience matrix EX1 may be represented in a hash table format according to
In step 930, a query vector QV1 may be formed. In case of a single query word Q1, the associated vector of the experience matrix EX1 may be used as the query vector QV1 as such. In case of a query comprising two or more query words Q1, Q2, the query vector QV1 may be formed as a linear combination of the vectors identified in step 930. In particular, the query vector QV1 may be formed as a sum of the vectors identified in step 930.
In step 940, the query vector QV1 determined in step 930 may be compared with the vectors of the experience matrix EX1. The comparing may comprise calculating a difference between the query vector QV1 and each vectors of the experience matrix EX1. When the experience matrix EX1 is represented in the hash table format, the non-zero values of the vectors may be specified by an array INDH. In step 950, one or more words associated with one or more vectors having the smallest difference to the query vector QV1 may be selected and provided as the prediction words. The prediction words may be used for controlling the operation of the system 500, as described above.
a shows generating one or more auxiliary words T2-T6 based on one or more original words T1. Using the auxiliary word instead of the original word or in addition the original word may sometimes be more useful when updating the experience matrix EX1 and/or when making a prediction.
The original word may be e.g. a string “2 Jan. 2012 09:03”. In principle, another substantially similar string “2 Jan. 2012 09:04” would be interpreted to a completely different word. This might reduce the efficiency of updating and using the experience matrix EX1. The aim of the updating operations might be e.g. to gather co-occurrence data, which allows predicting actions, which are likely to be executed in a certain day(s) of a week. For this aim, the original word “2 Jan. 2012 09:03” may be replaced or supplemented with an auxiliary word T5, which specifies the day of the week (Monday) corresponding to the date 2 Jan. 2012. The auxiliary word T5 may be determined from the original word e.g. by using a calendar.
The time expression contained in the original word “2 Jan. 2012 09:03” might have too high temporal resolution for some purposes. For example, if the aim is to predict what would be a suitable time to start e.g. a virus scanning application (which will consume battery power and data processing resources), an auxiliary word T6 may be determined such that the auxiliary word T6 represents the clock time with a lower resolution (e.g. “morning”) from the original word “2 Jan. 2012 09:03”. The auxiliary word may be provided e.g. by using a dictionary and/or by using a quantizer (classifier).
For example, the following words may be used to represent time:
time://late night (to represent times from 00:16 to 04:00)
time://early morning (to represent times from 04:01 to 06:00)
time://morning (to represent times from 06:01 to 11:45)
time://noon (to represent times from 11:46 to 12:15)
time://afternoon (to represent times from 12:16 to 17:00)
time://early evening (to represent times from 17:01 to 19:00)
time://evening (to represent times from 19:01 to 23:45)
time://midnight (to represent times from 23:46 to 0015)
The original word may specify e.g. location. The location may be specified e.g. by using GPS coordinates and/or by providing a name of a location. One or more auxiliary words specifying a location may be determined from the original word e.g. by using an electronic map. The map may be e.g. downloaded from the internet. The one or more auxiliary words may be provided by a remote (internet) service.
For example, a first device of the system 500 may be located in a place called as “Vantaa” (in Finland), and a second device of the system 500 may be located in a place called as “Espoo” (in Finland). For example, an auxiliary word “Southern Finland” or “Capital area of Finland” may be determined from the word “Vantaa” e.g. by using an electronic map or a location dictionary. For example, an auxiliary word “Southern Finland” or “Capital area of Finland” may be determined from the word “Espoo” e.g. by using an electronic map or a location dictionary. Thanks to the common auxiliary word, the co-occurrence data gathered by the first device may be used provide a prediction relevant to the user of the second device.
The user of the system 500 may also associate a location with a name. For example, a portable device carried by the user may determine its location e.g. by a GPS navigation unit, and the device may also ask the use to give a name for the detected location. The name given by the user may be subsequently used as the word associated with said (GPS) location. The name may be e.g. “home”, “school”, “university”, or “working place”. One or more auxiliary words T2, T3, T4 may be formed from the original word by tokenizing and/or parsing. The auxiliary word or words T2 to T6 may be used in a bag instead of or in addition to using the original word T1. Also a broader uniform resource locator (URL) may be determined based on a narrower uniform resource locator. In particular, a broader uniform resource locator of the internet may be determined based on a narrower uniform resource locator of the internet.
Referring to
Also an auxiliary word having a narrower semantic meaning (“species”) may be determined (by using a dictionary) based on an original word having a broader semantic meaning (“genus”). The system may be configured to determine a semantic descendant of an original word. This operation may be used for storing co-occurrence data e.g. in a situation where the original word is not contained in the vocabulary VOC1, but the term having the narrower semantic meaning is contained in the vocabulary VOC1. For example, a word “near 9 PM” may be generated based on an original word “evening”. However, this operation may sometimes involve a risk of adding partly erroneous information.
There may be a number of servers connected to the network, and in the example of
There are also a number of end-user devices such as mobile phones and smart phones 1251, Internet access devices (Internet tablets) 1250, personal computers 1260 of various sizes and formats, televisions and other viewing devices 1261, video decoders and players 1262, as well as video cameras 1263 and other encoders such as digital microphones for audio capture. These devices 1250, 1251, 1260, 1261, 1262 and 1263 can also be made of multiple parts. The various devices may be connected to the networks 1210 and 1220 via communication connections such as a fixed connection 1270, 1271, 1272 and 1280 to the internet, a wireless connection 1273 to the internet 1210, a fixed connection 1275 to the mobile network 1220, and a wireless connection 1278, 1279 and 1282 to the mobile network 1220. The connections 1271-1282 are implemented by means of communication interfaces at the respective ends of the communication connection. The various end-user devices may have one or more experience matrices or parts of experience matrices in their memory. The various end-user device may have capability of modifying the experience matrices and/or making a prediction using an experience matrix.
It needs to be understood that different embodiments allow different parts to be carried out in different elements. For example, execution of a software application may be carried out entirely in one user device like 1250, 1251 or 1260, or in one server device 1240, 1241, or 1242, or across multiple user devices 1250, 1251, 1260 or across multiple network devices 1240, 1241, or 1242, or across both user devices 1250, 1251, 1260 and network devices 1240, 1241, or 1242. For example, the capturing of user input through a user interface may happen in one device, the data processing and providing information to the user may happen in another device and the control and management of context processing may be carried out in a third device. As another example, forming an experience matrix may happen in one device (or a plurality of devices), processing the experience matrices may happen at one device, and making a prediction may happen at another device. The different application elements and libraries may be implemented as a software component residing on one device or distributed across several devices, as mentioned above, for example so that the devices form a so-called cloud. A user device 1250, 1251 or 1260 may also act as web service server, just like the various network devices 1240, 1241 and 1242. The functions of this web service server may be distributed across multiple devices, too.
The different embodiments may be implemented as software running on mobile devices and optionally on devices offering network-based services. The mobile devices may be equipped at least with a memory or multiple memories, one or more processors, display, keypad, motion detector hardware, and communication means such as 2G, 3G, WLAN, or other. The different devices may have hardware like a touch screen (single-touch or multi-touch) and means for positioning like network positioning or a global positioning system (GPS) module. There may be various applications on the devices such as a calendar application, a contacts application, a map application, a messaging application, a browser application, a gallery application, a video player application and various other applications for office and/or private use. The different devices may have various sensors e.g. for determining the physical properties or state of the environment or the devices such as temperature, time, humidity, lighting, sound volume, air pressure, position and movement.
The user device 1251 may comprise a user interface 501. A user may receive information via the interface 501. The user may control operation of the device 1251 and/or the system 500 by giving commands via the user interface 501. The user interface may comprise hardware, e.g. a display, keypad and/or a touch screen. The user interface may comprise a display screen for viewing graphical elements displayed on the screen. The user interface may also comprise a software application e.g. for displaying various different virtual keys on a touch screen.
A user interface 501 may be implemented e.g. in a stationary or movable device, which contains a camera arranged to receive commands from a user based on gesture recognition. The device may comprise a microphone arranged to receive commands from a user based on voice recognition. The user does not need to be in (mechanical) contact with a part of the user interface 501.
A user interface 501 may also be implemented in a device, which can be e.g. wrapped around the user's wrist (i.e. a wrist watch type device). A user interface 501 may be implemented e.g. in goggles (spectacles), which may comprise e.g. a virtual display for displaying information and/or a gaze direction detector to receive commands from a user by following his gaze direction.
The system 500 may be configured to carry out various tasks. Program code may be run in the system in order to carry out the tasks. These tasks may include e.g. creation of text, receiving and sending messages, playing music, capturing video, receiving news, receiving weather information, updating and viewing a calendar, browsing in the internet, and/or navigating in the streets of a city, etc. The system may further comprise one or more functional (physical) units to carry out the tasks. As the functional units, the system may comprise e.g. a video camera, a microphone, an audio reproduction unit, a display screen, an image projector, a radio frequency communication unit (WLAN, Bluetooth, NFC), a unit capable of communicating with a mobile communications network, a GPS navigation unit etc.
Words for the bags may obtained e.g. from one or more sensors G1. The sensors may include e.g. an acceleration sensor, a location sensor, a microphone, a radio receiver, an illumination sensor, a camera, an infrared radiation sensor, temperature sensor, air pressure sensor, magnetic compass.
A user device may be powered by a rechargeable battery. For example the user device may be carried by a traveling user such that the battery of the user device can be charged only occasionally (i.e. not at any time). Collected co-occurrence data may e.g. reveal in which kind of situations energy-consuming functionalities are typically activated by the user. Collected co-occurrence data may e.g. reveal when a possibility of recharging the battery is typically available.
In an embodiment, a battery operated system may be charged occasionally e.g. from an unreliable mains power network, from a local generator, from a solar cell and/or from a wind turbine. Co-occurrence data may reveal e.g. in which kind of situations energy-consuming functionalities are typically activated. Co-occurrence data may reveal how the availability of energy typically depends e.g. on the time (e.g. day or night), season (winter or summer), weather conditions (wind, non wind, sunlight, not sunlight). Co-occurrence data may reveal when a failure of the mains power network has typically happened.
The system 500 may comprise an uninterruptible power supply (UPS). In case of a power failure, unnecessary functionalities may be switched off based on predictions provided by the subsystem 400.
The word may be a string. The word may be a number. The word may comprise a sequence comprising letters and/or numbers. The word may comprise a sequence of ASCII codes represented in binary, decimal or hexadecimal format. The word may have a semantic meaning, but this is not necessary. The word may be a uniform resource identifier (URI) such as a uniform resource name (URN) or a uniform resource locator (URL). The word may be an abbreviation or an acronym. The “words” may also be called as “tags”.
The co-occurrence data is gathered to the experience matrix EX by using groups of words called as bags. A bag of words may also be called as a group of words or as a set of words.
A prediction can be made by using co-occurrence data stored in the experience matrix EX1. The “prediction” may also comprise determining a word representing a present event or a past event, i.e. it does not necessarily refer to a future event. The prediction may mean estimating the most probable word or words describing a situation.
The length of the vectors of the experience matrix EX1 does not need to be increased even when data is gathered from a high number of bags. In an embodiment, an experience matrix EX1 may comprise co-occurrence data gathered from NBG bags such that the number NBG of the bags (i.e. “documents”) is substantially higher than the number m of columns of the experience matrix EX1.
The “context” of a word may mean a verbal environment of said word such that said word (itself) is excluded from said environment. In this sense, a word does not belong to its own context.
The vectors of the experience matrix EX1 and the query vector QV1 may comprise information about the verbal environment of one or more words. The vectors of the experience matrix EX1 may also be called as context vectors. The query vector QV1 may also be called as a context vector.
In general, entries of the vocabulary VOC1 may be strings containing one or more symbols.
In an embodiment, entries of the vocabulary VOC1 may be numbers instead of being natural language words. Entries of an auxiliary table may be natural language words, and the auxiliary table may contain pointers to the numbers of the vocabulary VOC1. A word of natural language may be mapped to a number by using the auxiliary table. A number may be mapped to a word of natural language by using the auxiliary table. The auxiliary table may be called e.g. as a natural language dictionary. In particular, the auxiliary table may be a multilingual dictionary containing words of several different natural languages. A first word of a first natural language and a second word of a second natural language may be associated with the same entry of the vocabulary VOC1 by using the multilingual dictionary. A first word of a first natural language and a second word of a second natural language may be associated with the same number of the vocabulary VOC1 by using the multilingual dictionary. The multilingual dictionary may comprise e.g. an English-to-number dictionary, a Finnish-to-number dictionary and/or a German-to-number dictionary. Thus, the same vocabulary VOC1, the same basic matrix RM1 and/or the same experience matrix EX1 may be used with English, Finnish and/or German words.
The words of this table are not to be translated during proceedings of this patent or patent application.
Thus, the system 500 may handle multilingual information. The system 500 may gather information from several different states e.g. via the Internet. The resulting experience matrix EX1 may be language-independent. The same experience matrix EX1 may be used globally.
a and 12b show different data processing devices where forming and/or modifying and/or determining the context (or making the prediction) may be carried out according to an example embodiment. As shown in
a and 13b show various ways of combining two experience matrices or parts of experience matrices, and various resulting experience matrices (data structures).
According to a method A, two experience matrices 1 and 2 may be combined by adding the sparse vectors (denoted by horizontal bars) of experience matrix 2 for words (e.g. “Lapland, Rudolf, reindeer”) to the experience matrix 1, for example at the end, in the middle or by mixing in an interlaced manner. The resulting experience matrix comprises sparse vectors from matrix 1 such as for “pen” the vector “2 0 0 0 0 0 7 0 0 0” and for “putter” the vector “0 0 3 0 0 0 0 0 0 0”. For practical reasons, the vectors shown here are short and less sparse than the actual vectors used, as described earlier. The resulting matrix also comprises vectors from matrix 2 such as for “Rudolf” the vector “0 0 0 0 0 0 0 0 −7 0” and for “reindeer” the vector “0 −2 0 0 0 4 0 0 0 0”. The resulting matrix is such that when a closest match for an input vector is being sought, the search may return vectors/words from both the original matrix 1 and matrix 2. In one alternative, the number of words (and vectors) to pick from matrix 1 and from matrix 2 is received from a user through user interface e.g. using a slider control, or by determining the combination ratio in another manner.
According to a method B, two experience matrices 1 and 2 may be combined by adding elements of sparse vectors from experience matrix 2 to the sparse vectors of matrix 1 on a vector-by-vector basis, for example at the end, in the middle or by mixing in an interlaced manner. The resulting experience matrix comprises sparse vectors that have elements from matrix 1 such as for “pen” the vector elements “2 0 0 0 0 0 7 0 0 0” and for “putter” the vector elements “0 0 3 0 0 0 0 0 0 0”. The resulting matrix also comprises vector elements from matrix 2 such as for “pen” the vector elements “0 0” and for “putter” the vector elements “0 −7”. The resulting vectors are thus for “pen” the vector “2 0 0 0 0 0 7 0 0 0 0 0” and for “putter” the vector “0 0 3 0 0 0 0 0 0 0 0 −7”. That is, as an example, the corresponding vectors from matrix 1 and 2 may be concatenated. The resulting matrix is such that when a closest match for an input vector is being sought, the search may return words that would be a close match in either or both the original matrix 1 and matrix 2. In one alternative, the number of elements to pick from matrix 1 vectors and from matrix 2 vectors is received from a user through user interface e.g. using a slider control, or by determining the combination ratio in another manner.
According to a method C, two experience matrices 1 and 2 may be combined by performing the combination element by element, e.g. by summing the individual elements of corresponding vectors. The resulting matrix comprises sparse vectors that have elements that are combinations of elements from matrix 1 and matrix 2. For example, each element may be formed by summing, by subtraction, by taking a maximum or a minimum or by any other form of combining. As an example, “pen” has a corresponding vector “2 0 0 0 0 0 7 0 0 0” in experience matrix 1 and a corresponding vector “0 0 0 −7 0 0 0 0 0 0” in matrix 2, and the combined (in this case, summed) result vector for the word “pen” is “2 0 0 −7 0 0 7 0 0 0”. For “putter” the vectors from matrices 1 and 2 are “0 0 3 0 0 0 0 0 0 0” and “0 0 1 0 0 0 0 −5 0 0”, and the resulting vector for “putter” is “0 0 4 0 0 0 0 −5 0 0”. The combining may also happen in a weighted manner, e.g. so that the relative weighting is received from a user through user interface e.g. using a slider control, or by determining the weighting ratio in another manner.
The various combination methods may be used together as shown in
In using methods A and B together, vectors may be added, and the corresponding vectors may be combined. Vectors may also be modified so that some elements (the smaller rectangle inside the matrix) are taken from matrix 2 and some from matrix 1. In other words, the number of vectors in the result matrix may be different from the original number of vectors in either matrix 1 and 2, and the number of elements in each vector may be different from the original vectors.
In using methods A and C together, a vector in matrix 1 that has a corresponding vector in matrix 2 is combined element by element (as in a method C), and the vectors in matrix 2 that do not have a correspondence in matrix 1 are added to matrix 1.
In using methods B and C together, corresponding vectors from matrix 1 and matrix 2 may be combined, or parts of the corresponding vectors may be combined. For example, a first number of elements may be taken from matrix 1 and a second number of elements may be taken from matrix 2. In addition to combining by e.g. concatenation, the vectors may also be combined by element-by-element combination, which may happen in a weighted manner.
In using methods A, B and C together, different sizes of experience matrices may be combined. For example, corresponding vectors (vectors corresponding to the same word) may be combined by adding the corresponding elements on an element-by-element basis, and concatenating the elements from the longer vector. In addition, those vectors that do not have a corresponding vector, are added to the result matrix. The result matrix may thus comprise vectors for all words that have a vector in either experience matrix 1 or 2, and the vectors may be a combination of vectors in case a corresponding vector exists in both experience matrix 1 and 2.
In all methods, the combining may happen in a weighted manner, e.g. so that the relative weighting is received from a user through user interface e.g. using a slider control, or by determining the weighting ratio in another manner. The weighting may control e.g. the weighting in a weighted sum in the element-by-element combination, or the (relative) number of elements to be taken from matrix 1 and matrix 2. The weighting may be linear in the sense that element values multiplied by weighting coefficients may be taken from matrix 1 and matrix 2 to form the combined matrix. Alternatively, the weighting or combination may happen in a non-linear manner e.g. so that values from matrix 1 and matrix 2 are mapped non-linearly to mapped values 1 and 2, and these are then combined. The combining may also take place so that the combined element value c is obtained as a function from matrix 1 and 2 element values a and b, respectively, through a mapping c=f(a,b), where f is the mapping function. The combined values may be integer values, e.g. obtained by rounding, or they may be rational or irrational values, e.g. represented by floating point representation. The weighting may also take place so that some vectors or some elements are boosted e.g. by multiplying their element values with a coefficient (e.g. larger than 1.0) before combining.
The combination may also happen as a subtraction. For example, an earlier version of an experience matrix may be subtracted from a later version of the same experience matrix. The experience matrix may have evolved e.g. through training. In such a case, the result matrix would contain purely the change between the two matrices. In such a manner, matrices may be formed that can be added to a “baseline” experience matrix in case certain information needs to be available e.g. for making predictions in a certain situation or certain context.
It needs to be understood that more than two experience matrices (e.g. three, four, five or more) may be combined in a similar manner as presented above. That is, the combining described above is not limited to two experience matrices. In fact, even a large number of experience matrices may be combined in this manner, e.g. 20, 50 or 100 or even more, e.g. related to a social network service and e.g. experience matrices of a user and all his friends.
In phase 1420, at least a part of at least one sparse vector of the combined experience matrix is accessed to form a prediction output. The resulting combined matrix may be stored in the same apparatus that accesses the matrix, or at another device, for example in a network service. The accessing may comprise e.g. determining a close match for a sparse vector, or otherwise determining corresponding sparse vector(s) in the combined experience matrix. The forming of the prediction output may happen similarly as explained earlier for a general experience matrix.
In phase 1430, a system may be controlled or adapted in response to or by using the prediction output from phase 1420. For example, the user interface of a system may be altered or controlled by using the prediction output. The user interface may be that of an apparatus or that of a web service. A system may also be controlled or adapted in another manner, e.g. prediction output may be used to recommend websites (addresses) to the user so that the user may find relevant information from the websites in the current context. A system may also be adapted so that it finds or recommends relevant applications to be downloaded or installed onto the system. The recommending of websites or applications may happen at a user device or at a web service (server device).
The experience matrices may be combined using equal weight but they can also be combined with different weight values. This can for example happen so that user is presented a user Interface (UI) where user has a bar that he/she can move from right to left. Leftmost position may equal to giving all weight to user's own matrix X which would result in all predictions to be fully personal for that user. Rightmost position could give all weight to other matrix Y which would result in all predictions to be fully matrix Y specific, that is, based on the social experience matrix. Middle point would take both matrixes equally in.
With this kind of system user would have control how much he/she wants to give weight to his/hers personal information and how much he/she wants to rely on ‘social’ information.
A cloud based system may be used to collect experience matrixes from individual users. In this implementation each client device may send its own experience matrix to cloud. The cloud then creates a combined experience matrix as explained earlier. After this combined matrix has been formed it may be send to client devices.
Matrix sharing may happen through other channels as well like using proximity radios like Bluetooth or NFC. In these cases sharing of the experience matrix could take place directly between client devices.
After a client device receives an experience matrix from the cloud it may store it locally on device. Device can offer a UI for end user that allows user to configure how predictions are constructed on the device. In this case relevant parameters are how much system uses device owner's personal experience matrix and how much it uses the social matrix. This configuration may be offered to end user for example using a user interface.
Also, the user can periodically store a back-up of his or her random matrix. The back-up may be stored in a way that contains only the incremental changes from certain period as explained earlier. A comprehensive user history can be created by combining all the periods. A combination of the user comprehensive history with the latest period can be used for adopting changes in user patterns, when detected.
The reweighting of the different instantiations of gathered random matrices may be triggered from a change in user patterns. These can be detected from the worsened prediction accuracy of the system (i.e. the user is not choosing the proposed items).
For example, during holiday season and vacations the relevant experience matrices should be gathered from those time periods that the user has been vacating. This allows also the ability to share someone else's vacation random matrix component and then remove it afterwards, like describes previously in using the “social” random matrix.
Also, we may assume that these matrices may become purchasable items, from celebrities or done just in purpose to create, for example, a great holiday experience.
In phase 1520, at least the second matrix may be received from another device or system over a communications connection. For example, the first experience matrix may be an experience matrix of a user, and the second experience matrix may be a social experience matrix having been formed by combining at least a plurality of other experience matrices than the first matrix, for example experience matrices of people in a social relationship with the user in a social network service. The first matrix may or may not be in the group of matrices to be combined, that is, the social experience matrix may comprise information of the current user's experience matrix, or it may not.
In phase 1525, a method and parameters for combining the first and second matrix may be determined. For example, combining the first experience matrix and the second experience matrix may be performed according to a combination weight indicating the respective weight of said first and second experience matrices in said combined experience matrix. The weight may be determined e.g. by receiving it from a user e.g. using a slide control on a user interface.
In phase 1530, a selection from a user may be received for selecting at least one experience matrix, so that the combining may be carried out based on the selection. For example, the user may select topics for which experience matrices are to be combined, and/or persons whose experience matrices are to be used. The selection may be received from a user apparatus at a network service. Based on the selection, the experience matrices to be combined may be determined based on the user selection. Then, at least a part of at least one experience matrix may be accessed or received based on the selection for combining experience matrices.
In phase 1535, at least a part of a first experience matrix and at least a part of at least a second experience matrix is caused to be combined. The combination may result in a combined experience matrix, having information from both the original matrices. The combination may happen as explained with
In phase 1540, an incremental experience matrix corresponding to another experience matrix may be formed. The incremental experience matrix may comprise special data of at least one system or apparatus. The special data may be related to a restricted context, for example in the sense of time or topic, or the people that the experience matrix relates to. For example, an incremental experience matrix may be formed using a backup or snapshot of an experience matrix, for example by subtracting the backup or snapshot experience matrix from a later experience matrix, the later experience matrix being a different version of said backup or snapshot experience matrix.
In phase 1545, the combined experience matrix may be received or sent over a communications connection to be used in another system or apparatus than where said combined experience matrix was formed.
In phase 1550, the combined experience matrix may be provided to be accessible at a communication interface, e.g. in a network service so that it can be downloaded or accessed in another way.
In phase 1555, at least a part of at least one sparse vector of the combined experience matrix is accessed to form a prediction output. The resulting combined matrix may be stored in the same apparatus that accesses the matrix, or at another device, for example in a network service. The accessing may comprise e.g. determining a close match for a sparse vector, or otherwise determining corresponding sparse vector(s) in the combined experience matrix. The forming of the prediction output may happen similarly as explained earlier for a general experience matrix. The accessing may take place across a communications connection.
For example, a set of words describing at least a current context of a user system may be formed, for example describing previously and currently visited web sites. The combined experience matrix may be used for determining at least one prediction of a web address relevant to the set of words. As another example, a set of words describing at least a current context of a user system may be formed, describing for example applications residing in the system. The combined experience matrix may be used for determining at least one prediction of an executable application relevant to the set of words.
In phase 1560, a system may be controlled or adapted in response to or by using the prediction output from phase 1555. For example, the user interface of a system may be altered or controlled by using the prediction output. The user interface may be that of an apparatus or that of a web service. A system may also be controlled or adapted in another manner, e.g. prediction output may be used to recommend websites (addresses) to the user so that the user may find relevant information from the websites in the current context. A system may also be adapted so that it finds or recommends relevant applications to be downloaded or installed onto the system. The recommending of websites or applications may happen at a user device or at a web service (server device).
For example, at least one web address may be provided to a user for example as a recommendation by displaying recommended web sites. In addition or instead, a search engine search may be performed or created in web content based on the prediction of a web address. As another example, at least one prediction of an application may be provided to a user for example as a recommendation by displaying recommended applications. Accessing or providing access to downloading or purchasing an application from an application store may be made possible or performed based on said prediction of an executable application.
In phase 1570, an experience matrix may be uploaded to app store or downloading an experience matrix from an app store.
An experience matrix based prediction may be used for recommending a web address. In this embodiment, a user may receive news, feeds and advertising suggestions based on user's past activity, location and time. Basically an aspect that is supported as a word in an experience matrix may be used as a parameter when deciding what is relevant for a particular user. Data that is stored into an experience matrix may be used to filter relevant information. The user may not need search or go through existing feeds sent to him/her to find what he/she wants. In addition user may not need to manually enter or input anything. Relevant words may be extracted from the experience matrix based on the given context.
Based on contextual information (location, time, sensors, previous visited pages etc.) a random index based algorithm calculates the most relevant words for the user in the given context as explained earlier with
As these news and feeds have been filtered using words that are relevant for end user most of the news and feeds should be relevant for end user in the context where he/she is.
A client device may take care of collecting and maintaining the context data in the experience matrix. Client device may be able to extract most relevant words for a given context from the experience matrix. This list of keywords may be based for example on sensor data (such as movement), time of the day, location etc.
After the client has collected relevant words it may send the words to server that is located in the network. A network server is used as it may have more bandwidth and processing power. A network server may collect news and social feeds from a variety of different network sources like twitter, facebook, bbc.com, cnn.com etc.
When a server receives a word list from the client it may start filtering the news and social feeds based on these words. It may select the news and feeds so that they provide best fit to the word list provided by the client. After the server has completed filtering it may return a list of filtered news/social feeds to client. Or alternatively it can store results to a specific location from where client can fetch the results.
Similarly, the user can add/remove key words on his/her own. There the user makes the desired change in the UI and client sends that change to the server which interprets it as one of the key words. The user fed key words may have more weight in creating the recommendations or they can be handled as other, system created key words. The search and recommendation may be given back in real time or the result may be given when next time in the context.
An experience matrix based prediction may be used for recommending a downloadable or installable application. In this embodiment, a user may receive application suggestions based on user's past activity, location and time. Past activity also includes the usage activity of existing applications. An aspect that is supported as a word in an experience matrix may be used as a parameter when deciding what is relevant for a particular user. Data that is stored into an experience matrix may be used to filter relevant information. The user may not need to search or manually enter or input anything. Relevant words may be extracted from the experience matrix based on the given context.
Based on contextual information (location, time, sensors, previous used applications etc.) a random index based algorithm may predict the most relevant words for the user in the given context. The predicted words may be used as filtering words for a search at the server side. The client device may send these words to network server. The server may go through the extracts or descriptions of applications from one or more application stores in question and other Internet sources (such as application review sites). Based on words received from the client device the server will filter apps and select a limited number of applications or links to applications that will be delivered to client.
As these applications have been filtered using words that are relevant for end user, the applications may be more relevant for end user in the context where he/she is. Also the application suggestions may come actively to the user, without the user having the need to check what interesting is available. Alternatively, the system may also install these applications automatically without any user intervention.
The word extractions and finding the relevant applications may happen e.g. as described above for the web addresses.
In addition, the user can influence in settings if she/he would like to prioritize applications with higher rankings, or ones which friends have in their experience matrices. Also, a user can enable the system to automatically install the selected applications.
The various embodiments of the invention can be implemented with the help of computer program code that resides in a memory and causes the relevant apparatuses to carry out the invention. For example, a terminal device may comprise circuitry and electronics for handling, receiving and transmitting data, computer program code in a memory, and a processor that, when running the computer program code, causes the terminal device to carry out the features of an embodiment. Yet further, a network device may comprise circuitry and electronics for handling, receiving and transmitting data, computer program code in a memory, and a processor that, when running the computer program code, causes the network device to carry out the features of an embodiment. A data structure may be formed so that when the data structure is used in a computer with a computer program for controlling the same, the data structure causes the computer program and the computer to operate according to the data structure and the underlying data.
It is obvious that the present invention is not limited solely to the above-presented embodiments, but it can be modified within the scope of the appended claims.
| Filing Document | Filing Date | Country | Kind | 371c Date |
|---|---|---|---|---|
| PCT/FI2012/050179 | 2/22/2012 | WO | 00 | 12/1/2014 |
