Learning-based strategies for message-initiated constraint-based routing

Abstract

A method is presented for a learning-based strategy utilized within message-initiated constraint-based routing for digital message communication among nodes in an ad-hoc network, in which each node includes attributes. The method includes determining local attributes for each of the nodes and defining constraints on the attributes. Each node is provided access to the attributes of each neighboring node. Each message transmitted over the network has a message type, which includes a destination specification, route specification, and objective specification. Constraint checking and cost estimation checking are performed for each message type. Cost estimation is utilized to converge on an optimal message path. The message that is routed within the network includes the address of a sending node, the address of the source node, route constraints, destination constraints, the number of route constraints, the number of destination constraints, message identification number, sequence identification number, and routing objectives.

Description

INCORPORATION BY REFERENCE

[0003] The following U.S. patents are fully incorporated herein by reference: U.S. Pat. No. 6,304,556 (“Routing and Mobility Management Protocols for Ad-Hoc Networks”); and U.S. Patent No. 5,570,084 (“Method of Loose Source Routing over Disparate Network Types in a Packet Communication Network”).

BACKGROUND OF THE INVENTION

[0004] This invention relates generally to communication protocols which are particularly suitable for self-reconfigurable multi-purpose communication networks, such as ad-hoc networks. More particularly, the protocol utilizes learning-based strategies to achieve routing objectives.

[0005] Various routing mechanisms have been proposed for ad-hoc wireless networks. In general, an ad-hoc wireless sensor network has the following properties: (1) the structure of the network is unknown and may change dynamically, (2) each node has limited computation resources and lifetime, and (3) each node can obtain pieces of information from local sensors and communicate with others within a limited range. The power of such sensor networks is derived from communication, since each node is only able to sense local information with little computational resources. The routing mechanisms proposed for such networks fall into two basic categories, table-driven or source-initiated. Table-driven protocols rely on an underlying global routing table update mechanism for all nodes in the network, a mechanism that would not be energy efficient for ad-hoc dynamic networks. Source-initiated protocols, on the other hand, discover a route every time it is needed.

[0006] Existing routing protocols differ mainly in routing metrics, but all use a fixed routing objective. In most cases, routing objectives are implicitly embedded in strategies. Examples of these routing metrics include use of the shortest path, degree of association stability, signal stability or strength combined with shortest path, and information gain. Protocols also differ by destination specifications. The majority of early protocols are address-based or geographical location-based.

[0007] All existing routing protocols for wireless networks are implicitly associated with their routing strategies, which generally fall into two classes, flooding-based or search-based. Flooding-based methods begin with a route discovery phase (flooding the network), followed by a route maintenance phase for repairing disconnected routes. Flooding-based strategies are more suitable for relatively stable networks, since maintaining and repairing routes can be costly for dynamic networks. Search-based methods normally discover routes by selecting the next “best” hop at every node on the route. Routes may differ from message to message, even to the same destination node, and there is no route maintenance.

[0008] However, existing ad-hoc wireless protocols do not have theoretical results on delivery or route optimality. Distributed quality-of-service routing for ad-hoc networks has been proposed, in which a set of probes is used to find an optimal route before actual messages are sent, followed by a route maintenance phase to repair the broken route. But this approach is not suitable for dynamic networks in which there is no fixed optimal route over time.

[0009] Existing routing mechanisms for ad-hoc wireless networks have two limitations: routing objectives are fixed and embedded in strategies and quality-of-service routing does not work well for dynamic networks. It would be useful to have a framework of distributed routing strategies based on real-time reinforcement learning, so that messages discover and learn their routes on the way to their destinations. The separation of routing objectives and routing strategies would also make it possible for network systems to change routing objectives from time to time, given different task characterizations and requirements.

SUMMARY OF THE INVENTION

[0010] Briefly stated, and in accordance with one aspect of the present invention, a method is presented for a learning-based strategy utilized within message-initiated constraint-based routing for digital message communication among nodes in an ad-hoc network, in which each node includes attributes. The method includes determining local attributes for each of the nodes and defining constraints on the attributes. Each node is provided access to the attributes of each neighboring node, with a neighboring node being a node that is one hop away. Each message transmitted over the network has a message type, which includes a destination specification, route specification, and objective specification. Constraint checking and cost estimation checking are performed for each message type. Cost estimation is utilized to converge on an optimal message path. The message that is routed within the network includes the address of a sending node, the address of the source node, route constraints, destination constraints, the number of route constraints, the number of destination constraints, message identification number, sequence identification number, and routing objectives.

[0011] In accordance with another aspect of the invention, there is presented a system for a learning-based strategy utilized within message-initiated constraint-based routing for digital message communication among nodes in an ad-hoc network. The system includes a local attribute module for determining local attributes for each of the nodes in the ad-hoc network. A remote attribute module provides access to the attributes of each neighboring node, with a neighboring node being a node one hop away from a current node within the ad-hoc network. A timer module provides a time trigger function for the local attribute module and the remote attribute module. A broadcast module provides a send function for the local attribute module and a receive function for the remote attribute module. An attribute property module estimates the minimum and maximum values of at least one attribute. A constraint module defines constraints on the attributes, performing constraint checking for each message type, in which a message type includes a destination specification, route specification, and objective specification, and performs cost estimation checking for each message type. A routing module for routing a message within the ad-hoc network utilizes a learning-based strategy to converge on an optimal message path, with the message including the address of a sending node, the address of the source node, route constraints, destination constraints, the number of route constraints, the number of destination constraints, message identification number, sequence identification number, and routing objectives.

[0012] In accordance with yet another aspect of the invention, there is presented an article of manufacture in the form of a computer usable medium having computer readable program code embodied in the medium which causes the computer to perform method steps for a learning-based strategy utilized within message-initiated constraint-based routing for digital message communication among nodes in an ad-hoc network. The method includes determining local attributes for each of the nodes and defining constraints on the attributes. Each node is provided access to the attributes of each neighboring node, with a neighboring node being a node that is one hop away. Each message transmitted over the network has a message type, which includes a destination specification, route specification, and objective specification. Constraint checking and cost estimation checking are performed for each message type. Cost estimation is utilized to converge on an optimal message path. The message that is routed within the network includes the address of a sending node, the address of the source node, route constraints, destination constraints, the number of route constraints, the number of destination constraints, message identification number, sequence identification number, and routing objectives.

BRIEF DESCRIPTION OF THE DRAWINGS

[0013] The foregoing and other features of the instant invention will be apparent and easily understood from a further reading of the specification, claims and by reference to the accompanying drawings in which:

[0014] FIG. 1 is a schematic illustration of a communication network comprised of a plurality of nodes which communicate with one another in accordance with a routing scheme that includes an embodiment of the present invention;

[0015] FIG. 2 is a diagram of the components in a forward message specification according to one embodiment of the present invention;

[0016] FIG. 3 is a diagram of the components in a confirmation message specification according to one embodiment of the present invention;

[0017] FIG. 4 is a block diagram of one embodiment of the system showing modules which may be employed in accordance with the protocol specification of the present invention;

[0018] FIG. 5 is a flow chart illustrating an embodiment of the attributes management module;

[0019] FIG. 6 is a flow chart illustrating an embodiment of the remote attributes management module, in which neighbor's attributes are updated and neighbor's message count is incremented;

[0020] FIG. 7 is a flow chart illustrating an embodiment of the remote attributes management module shown, in which a neighbor's active state is adjusted;

[0021] FIG. 8 is a flow chart illustrating the use of a search based strategy with an embodiment of the present invention;

[0022] FIG. 9 is a flow chart illustrating the Q-learning confirmation block in accordance with an embodiment of the present invention;

[0023] FIG. 10 is a flow chart illustrating the Q-learning next hop selection block according to an embodiment of the present invention;

[0024] FIG. 11 is a flow chart illustrating the handling of confirmation MCBR messages according to an embodiment of the present invention;

[0025] FIG. 12 is a flow chart illustrating an embodiment of Q-learning back propagation according to an embodiment of the present invention.

DETAILED DESCRIPTION OF THE INVENTION

[0026] Although various routing strategies have been developed for ad-hoc wireless networks, they are usually tightly integrated with routing objectives. Furthermore, most of the existing routing strategies do not have theoretical guarantees on message delivery and route optimality. Quality of Service (QoS) routing does compute routes satisfying constraints and optimality, but in general they do not work well for dynamically changing ad-hoc networks. Maintaining a QoS route normally involves extra message exchange, which in turn results in additional energy consumption costs.

[0027] In contrast to this, the framework of distributed routing strategies described herein is based on real-time reinforcement learning. Such strategies are localized, in the sense that no global knowledge of the network is assumed. Unlike other QoS routing approaches, in which an optimal route must be found before the transmission of actual messages, with distributed routing strategies messages discover and learn their routes on their way to their destinations. No route repair and maintenance are necessary; therefore such strategies are more adaptive to dynamic changes of the network. This type of routing can be called “connectionless” routing, since there are no explicit connections between the source and the destination. However, like many other routing protocols, symmetric communication links ((v, w)∈E if and only if (w, v)∈E) are assumed for this type of strategy.

[0028] There are two goals for learning-based strategies. The first is to route a message eventually to a destination if a path exists, and the second is to route a message optimally with respect to its objective. Unlike most existing localized routing strategies, where no global optimality is defined, these strategies make use of the specified objectives and converge to optimal routes defined by messages if the network is stable. This type of routing may also be called “QoS-aware” routing. Furthermore, within this framework an optimal path is found from the source to a destination satisfying the destination constraint, while keeping the routing constraint satisfied on all intermediate nodes (unicast routing). If there is more than one destination, the messages may be delivered to different destinations at different times before converging to the optimal destination.

[0029] This framework is placed within a message-initiated constraint-based routing protocol, in which the routing destination(s), routing constraints and objectives, as well as routing strategies are explicitly specified. Message-initiated Constraint-based Routing (MCBR) as used herein describes routing mechanisms with constraint-based destinations and objectives specified in messages. In MCBR, each node in the network has a list of attributes, whose types are predefined and known globally. Attributes can be anything from geographical locations to network bandwidths, from sensor values to internal clocks. The values of attributes can be constant, such as a node identifier or a unit cost, or can change from time to time. For example, a mobile node may change its locations; a stationary node can still obtain different sensor readings although its environment changes. A routing destination is explicitly represented by a set of constraints on attributes. This destination specification is more general than attribute-based specification, since constraints may describe any relationship or characteristic. Furthermore, in addition to destinations, local route constraints, if any, are explicitly specified. Examples of local route constraints include: avoiding a noisy area, avoiding congestion, and avoiding low-energy nodes, etc. Finally, a routing objective is explicitly stated, such as a shortest path, maximizing energy levels over the route, maximizing connectivity over the route, or minimizing congestion, etc.

[0030] A portion of an example network 100 that includes a plurality of communications nodes 120 labeled A, B, C, D, E, and F is illustrated in FIG. 1. Each node has a unique address, which is encoded as a number or other symbolic representations known in the art. In this example network, the routing zone for each of the nodes 120 is defined by a corresponding one of a plurality of circles 140 whose radius equals the routing zone radius. Thus, for example, if node D broadcasts, each of the nodes within its routing zone 150 will receive the broadcast message, in this example, nodes A and E. Each node 120 includes a wireless transceiver and receiver, which communicates with neighboring nodes, and a routing module that passes messages from a source to one or a set of destination nodes. Additionally, each node may have a set of sensors that can collect a variety of local information. In this example, the arrows connecting nodes A, D. E. F, and B form one possible message transmittal path within the sample network. It will be appreciated that for a mobile ad-hoc network, the nodes and connections will change from time to time, due to mobility, failures, battery life, or power management.

[0031] Each node also includes attributes, which consist of a data entity having a type and a domain of values. An attribute value denotes the current value of an attribute. An attribute may be a constant, such as the node identifier or the unit energy cost for data transmission. An attribute may also be a clock that increases monotonically or a sensor reading from a light or temperature sensor, which may vary when the local environment changes. Additionally, attributes may be monitors of the node's conditions such as battery level or computations resources, or they may represent network properties such as radio strength, signal loss or reliability, connectivity with neighbors, number of routes through the node, etc. Attributes can be estimated values via calculations such as the node's geographical locations or a target's speed and direction, or they may be properties of nodes, such as being mobile or stationary, being group leader or group member, etc. Attributes may also be values passed through messages, such as the number of hops away from the source. Attribute values can be accessed via their types, with the set of attribute types predefined and known globally.

[0032] Each message sent on the network has its destination, which may be one node or a set of nodes, but existing protocols do not specify route constraints, i.e., nodes that a message should avoid while routing to its destination. In MCBR, both destinations and route constraints are specified in messages, as illustrated in FIG. 2. The MCBR forward message 200 includes various components such as the sender's address 210 and the source address 220. The sender's address 210 is necessary for forwarding link establishment or when back propagation from the destination to the source node is needed. Consequently, sender's address 210 changes, as it is updated at each node through which a message passes, but source address 220 remains constant. Each node may keep a list of entries, each of which corresponds to a type of message with a specific source, destination and route constraints. The entries are created by new messages and updated by subsequent messages of the same type.

[0033] Constraints 290 can be defined on attributes as a set of variables. Formally, a constraint C is a pair <R,r>, where R is a set of attributes and r is a relation defined on R. If |R| is n, r is an n-ary relation. The value of C is true, or C is satisfied, at a node v, if and only if the current value of the attribute tuple located at v is in r. A simple unary constraint is a range constraint l≦a≦u, where a is an attribute, l and u are lower and upper bounds, respectively. Attribute-based specification thus becomes constraints, which can be aggregated via Boolean operations. An aggregated constraint C is a Boolean function b defined on a set of constraints, b:B1×B2× . . . Bn→B, where Bi is the Boolean domain for constraint Ci. C is satisfied at node v if and only if the value of the Boolean function is true given the values of the constraints at node v. For example, if the Boolean function is logical and, the aggregated constraint is satisfied if and only if all the constraints are satisfied. If the Boolean function is logical or, the aggregated constraint is satisfied if and only if one of the constraints is satisfied. In the embodiment shown in FIG. 2, only logical and is used implicitly. However, it is possible to encode the whole Boolean function as well. The destination of a message can be specified by a constraint or an aggregated constraint. Given a destination constraint Cmd of message m, a node v is a destination node for m if and only if Cmd is satisfied at v. The set of destination nodes, denoted by Vmd, is called the destination for m. For example, address-based routing, i.e., sending a message to a node with an address ad, can be represented using the destination constraint a=ad where a is the address attribute. Geographical routing, i.e., sending a message to a geographical circular region centered at (x0,y0) within radius c can be represented using the destination constraint (x−x0)2+(y−y0)2≦c where x and y are location attributes. Sensor-based routing, for example, sending a message to hot nodes, can be represented using the destination constraint t≧tm, where t is the temperature attribute and tm is the minimum desired temperature. Constraints can be combined to refine the destination region, for example, sending a message to a hot node within a region, etc. The number of destination constraints for each forward message is specified at 250, with the aggregated constraints for the destination being the logical and of all the constraints, for this embodiment.

[0034] The number of route constraints is specified at 270, with the aggregated constraints for the route being the logical and of all the constraints, for this embodiment. A local route constraint extends the concept of failure in networked nodes, allowing a message to be routed only via a subset of nodes satisfying the constraint. Given a local route constraint Cmr of message m, the active network of <V, E> for m is a subnet <Vm, Em>, such that v∈Vm if and only if Cmr is satisfied at v and (v, w)∈Em if and only if v, w∈Vm and (v, w)∈E. For example, a message that should avoid nodes in light areas while routing to its destination has a local route constraint l≦lm where l is the light attribute and lm is the light intensity limit. High-priority messages and low-priority messages may be defined by different local route constraints: high-priority messages have no constraints, while low- priority messages will avoid nodes with high congestion.

[0035] An optional routing strategy identification (ID) number may be provided at 230. The distributed routing strategies described herein are characterized by the following properties: (1) localization: there is no global knowledge and there is no master computing the routing table, and (2) constant memory: each node has a constant memory that does not grow with the size of the path or the size of the network. Since different strategies may lead to different performance in different situations, the selection of a strategy can be made message-by-message. In the example embodiment, the strategy ID is specified as an 8-bit number, and each strategy ID is associated with a routine. The router dispatches to the associated routine (which may correspond to a particular strategy) according to the strategy ID.

[0036] The learning-based distributed routing strategy applies real-time reinforcement learning to distributed routing. (Real-time reinforcement learning is described more fully in “Reinforcement Learning: An Introduction”, R. S. Sutton and A. G. Barto, editors, the MIT Press, Cambridge, Md., 1998, hereby incorporated by reference in its entirety.)

[0037] Given a message m with MCBR specification <vm0, Cmd, Cmr, Om>, the corresponding network is <Vm, Em> and the MDP is <Vm, Em, f, r>, where

f (v, (v, w))=w,

r (v, (v, w))=om(w)

[0038] and om is the local objective function for m. The distributed constraint-based unicast routing is to find an optimal policy for m: πm: Vm→Em so that U(v)=Σt=1nom (vi) is minimized for all v∈Vm, where VN is a destination node satisfying Cmd. In other words, the distributed unicast routing is an undiscounted and deterministic reinforcement learning.

[0039] In order to implement the reinforcement learning algorithm, two assumptions are made: (1) the network has to be symmetric, i.e., if (v, w)∈E, then (w, v)∈E, and (2) each node has to store the list of attributes of its neighbors, so the actual set of neighbors with respect to the route constraints of a message can be determined. Each node broadcasts its attributes in a regular interval and broadcasts whenever an attribute changes significantly. This guarantees each node has a current up to date list of its neighbors and their attribute values. Regular broadcasting is not considered as overhead, since any search-based strategy has to keep an up-to-date list of neighbors.

[0040] Each node also maintains a list of entries, one for each type of message. For any implementation, the maximum number of entries is finite; an old entry can be reused by a new message, meaning the old estimation will be forgotten. Since the network is dynamic, retaining an entry for an extended period of time may not be meaningful. The entry is created the first time a new message type arrives. Each entry of type m at node v holds a list of estimation values, one for each of its neighbors, Qm(v, w), indicating the best estimate so far of the global objective om from its neighbor w to the destination. The estimates are set initially when an entry is created and updated by learning phases. The initial estimate is calculated according to heuristics derived from the destination and objective specification.

[0041] Q-learning-based routing consists of two phases: forward phase and backward phase. The forward phase is a policy improvement phase, i.e., deciding which neighbor to pass the message to according to its current estimation. The backward phase is a penalty learning phase; the estimation improves as more nodes are visited by this type of message. The backward phase is also used for message confirmation, so that if the confirmation has not arrived within a certain time period, the forward phase will be activated again for sending the message to a different neighbor. This guarantees that the message arrives at its next hop, if there is one. One example of pseudo code for this embodiment is provided hereinbelow for illustrative purposes. It will be appreciated that the various embodiments of the invention disclosed herein could assume numerous expressions in such code, all of which are contemplated by the scope of the specification and claims herein. In the example pseudo code, om (w) is the value of the local objective function Om of at node w, and Qm0 is an initial estimate of Qm. The algorithm does not assume a static network, i.e., <Vm, Em> can change from time to time. Even with a static network, the routes may vary from time to time as a result of learning. If Vmd is not singleton, it is also possible to deliver messages to different destination nodes at different times. To ensure a single destination, additional destination constraints can be added, for example, a unique node identifier.

Forward phase:
received message ( m) at node v from node u do
if new( m) then
for all w with (v, w) ∈ Em do
Qm(v, w)  Qm0 (v, w); end
end
send_Q(om(v) + minw Qm(v, w)) to u;
if satisfied(Cmd) then return; end
w  argminwQm(v, w); (random tie break)
send_message(m) to w;
end
Backward phase:
received_Q (Qm) at node v from node w do
Qm(v, w)  Qm(v, w) + α( Qm − Qm(v, w) );
end

[0042] For distributed routing approaches, two types of complexity can be defined. Message complexity measures the total number of messages sent in the network. Time complexity measures the maximum number of hops from the source to a destination node. There are two theoretical results regarding the complexity of a learning-based routing. One is the complexity for delivery of a message to its destination, the other is the complexity of optimal policy convergence. Learning-based routing does not guarantee the optimal policy for every message, however, it does guarantee that a message arrives at its destination if a path exists. Furthermore, if the network is stable, and there are many messages of the same type, it will converge to an optimal policy for that type of message.

[0043] If the network is assumed to be quasi-static, it does not change during the routing of one message. For this case, let Um(v)=minwQm(v, w). The value of Qm(v, w) is called admissible or under-estimated if

0<Qm(v, w)≦om (w)+Um* (w)

[0044] where Um* (w) is the optimal value of Om from node w to the destination. It is easy to see that: if Qm0 is admissible, then Qm stays admissible given the update rule in the routing strategy. Furthermore, Qm(v, w)<ommaxd, where ommax is the maximum value of the local objective function om, and d is the distance from v to the destination.

[0045] Message ID number 240 is specified as a number which corresponds to a unique type of the messages sent from its source. In this embodiment a message type includes the message components of the destination constraints, the route constraints and the objective. Similarly, sequence ID number 260 is specified as a number. The sequence ID number has two functions: (1) for flooding-based strategies, sequence number can be used to determine if the same message has been handled or not; (2) generally, sequence ID may be used to determine whether a message is lost or not, and also to assemble a series of messages to a large message according to its sequence ID.

[0046] Existing protocols, other than quality of service routing, do not explicitly specify routing objectives, which are implicitly embedded in routing strategies. MCBR explicitly specifies routing objective 280. To accomplish this, a local objective function o is defined on a set of attributes: o: A1×A2× . . . ×An→R+, where Ai is the domain of attribute i and R+ is the set of positive real numbers. The value of o at a node v, denoted o(v), is o(a1,a2 , . . . , an), where ai is the current attribute value of attribute i at node v. A local objective function can be a constant such as the unit energy cost. Multi-objectives can be obtained by a weighted sum of individual objectives, where the weights indicate the relative importance of individual objectives. For example,

O (v)=αo1(v)+(1−α)o2(v),

[0047] where o1 and o2 are local objective functions, and 0<α<1.

[0048] A local objective can be aggregated over the routing path to form a global route objective. There are two types of global objectives: additive or concave. A global objective function O of a local objective function o over a path p consisting of a sequence of nodes v0, . . . , vn is additive if 1$O\left(p\right)=\sum _{i=0}^{n}o\left(v_i\right);$

[0049] O is concave if 2$O\left(p\right)=\underset{i=0}{\overset{n}{\min }}o\left(v_i\right).$

[0050] For example, “shortest path” is a global additive objective defined on the constant local objective, one hop cost. The objective of “energy distribution” can be defined explicitly as follows. If the current energy level e is an attribute, the function au indicating used energy can be defined as a unary function u(e)=emax−e, and the global additive objective on local objective u, U(p)=Σi=0nu(vi), represents the “energy distribution” metric, that is, preferring routes with more energy. As another example, the bandwidth of a path can be represented as a concave objective B(p)=mini=0nb(vi) where b(v) represents the bandwidth of node v.

[0051] These two aggregation types are general. For example, convex aggregation 3$O\left(p\right)=\underset{i=0}{\overset{n}{\max }}o\left(v_i\right)$

[0052] can be represented by concave aggregations as 4$O\left(p\right)=\underset{i=0}{\overset{n}{\min }}\left(o^{\max }-o\left(v_i\right)\right);$

[0053] Multiplicative aggregation can be represented by additive aggregation since

log(Πo(i))=Σlog(o(i)).

[0054] In this example embodiment, only one additive objective is specified for the purposes of illustration.

[0055] In MCBR, received messages may be confirmed, as illustrated in FIG. 3. Each confirmation message 300 from the destination back to the source includes various components, such as the sender's address 310, which has the same meaning as in the forward message. However, sender's address 310 is changed to the current node address, rather than remaining as the previous node address, while source address 320 remains the same as the forward message. Message ID number 330 is the same as specified in the forward message and identifies each message with its unique message type. Cost estimate 340 may be encoded as a number and provides the estimation of the objective cost from the current node to the destination of the message, where the objective is specified in the corresponding forward message.

[0056] Turning now to FIGS. 4 through 7, there is shown a block diagram of one embodiment of the system for protocol specification for message-initiated constraint-based routing. FIG. 4 shows a component diagram of the current embodiment, where each component is depicted as a box and each arrow, from A to B, indicates that component A provides a set of functions for component B. In system 400, Timer Module 430 provides the time trigger function for both Local Attribute Module 420 and Remote Attribute Module 450. Broadcast Module 470 provides a send function for Local Attribute Module 420 and a receive function for Remote Attribute Module 450. Local Attribute Module 420 provides local attribute access for Attribute Property Module 410, Constraint Module 440 and MCBR Routing Module 460. Remote Attribute Module 450 provides neighbor attribute access for Attribute Property Module 410, Constraint Module 440, and MCBR Routing Module 460. As shown in FIG. 5, local attribute module 420 reads or computes local attributes at 510 when triggered by update timer 540, which is included in timer module 430 in FIG. 4. At 520 a determination is made as to whether there have been any significant changes to the local attributes. If significant changes have appeared, or when triggered by broadcast timer 550, which is included in timer module 430 in FIG. 4, local attributes are broadcast to its neighbors at 530. If significant changes have not appeared, local attributes are not broadcast, as shown at 560. Remote attribute module 450 in FIG. 4 is illustrated in more detail in FIGS. 6 and 7.

[0057] In FIG. 6, a node in the system receives remote attributes at 610. The neighboring node, which has provided the remote attributes, is identified from the sender's address at 620. The receiving node then updates the neighbor's attributes within its database and increases the neighbor's message count by 1 at 630. The message count is used to record how many messages are received between the triggers of Expire timer 750. As shown in FIG. 7, whenever the Expire timer triggers, for each neighbor 710 the node checks to determine at 720 whether there are enough messages received since the last trigger. If the message count is less than the threshold, then the neighbor is set to be inactive at 730. In either case, the message count is reset to 0 at 740. The message count review is triggered by expire timer 750, which is provided by timer module 430 of FIG. 4.

[0058] Returning to FIG. 4, attribute property module 410 provides the minimum and maximum value and gradient estimation of an attribute within the neighborhood, which is used by constraint module 440. Constraint module 440 provides constraint checking and cost estimation given message type and provides this information to MCBR Routing Module 460. The function of Routing Module 460 is explained further in FIGS. 8 and 11.

[0059] Because MCBR separates routing objectives from routing strategies, various generic strategies may coexist. Since different strategies may lead to different performance in different situations, the selection of a strategy can be made message by message. Turning now to FIG. 8, there is illustrated a flow chart of a search-based strategy 800 utilized with MCBR. Generally, search-based methods use greedy algorithms or a distributed real-time search to establish a path. Search-based methods may be more energy efficient, since messages are forwarded on a single route (rather than being broadcast at every station, as with flooding-based strategies). In this embodiment, initially an MCBR message “m” is received from an identified node “j” at 810. A determination is made at 820 as to whether this is a new type message. The message type includes the destination specification, route specification, and objective specification. A new type message is one not matching message types already on record. If message “m” is a new type message, at 830 an entry for the message is created; if message “m” is not a new type message, at 840 its entry is located. At 850 the cost for message “m” is estimated and a confirmation is either sent back to node “j” or broadcast locally to all its neighboring nodes. Cost may be defined as any attribute, for example sound, light, etc.

[0060] The cost estimation operation 850 is illustrated more fully in FIG. 9. It is possible to derive that the overhead of delivering a message is bounded by Σv∈Vm Σw(Qm*(v, w)−Qm0(v, w)). While making Qm0 admissible is trivial, since Qm0 can be set to 0, making admissible initial estimates other than zero is not trivial. However, a good initial objective estimation for Qm0 can reduce overhead significantly.

[0061] The initial heuristics are based on both the properties of destination constraints and objective functions. Destination constraints are used to estimate the minimum number of hops to the destination, while objective functions are used to estimate the minimum cost along the path. Let C be a constraint and s(C) be the degree of satisfaction of C; s(C) is zero if and only if C is satisfied. Let ΔC be the maximum change possible for s(C) in one hop. The minimum number of hops based on constraint C is h(C)=s(C)/ΔC. Given the destination constraint as a Boolean formula of logical and's and logical or's, the estimation can be done recursively as follows. If C=C1{circumflex over ( )}C2, then h(C)=max(h(C1),h(C2)). If C=C1{circumflex over ( )}C2, then h(C)=min(h(C1),h(C2)). For example, if destination constraints are geographical,

C : (x=xd){circumflex over ( )}(y=yd),

[0062] and the maximum radio radius is R, given the current location (x,y), we can obtain

h (C)=max(|x−xd|/R, y−yd|/R).

[0063] Two constraints C1 and C2 are called mutually exclusive if, for any single hop, at most one of s(C1) and s(C2) changes. If C=C1{circumflex over ( )}C2, where C1 and C2 are mutually exclusive, then h(C)=h(C1)+h(C2), which would be a better estimate than max(h(C1),h(C2)). Given the estimate of the number of hops h to the destination, an additive objective can be estimated as ominh, where omin is the minimum value of o over the network.

[0064] Returning to FIG. 9, at 910 the total cost from the current node is estimated. This may be performed with

Q (n)=min(C+Q(i)),

[0065] where C is local cost and Q(i) represents the estimated cost for the total path from the neighbor node and Q(n) represents the total cost from node n. At 920 the estimated cost is returned to the previous node, node j. If the initial estimates are admissible, the learning-based strategy guarantees delivery if a path exists. It also guarantees convergence to an optimal route if the change rate of the network is slower than the convergence rate. Alternatively, rather than returning the estimated cost to the source node at 920, the estimated cost may be sent to all neighboring nodes.

[0066] Returning now to FIG. 8, a determination is made at 860 as to whether the destination has been reached. If the destination has been reached, the message will be handled at 880, or, optionally, a destination confirmation may be sent at 865 before the message is handled at 880. A received message is handled, if the routine corresponding to the receiving of that message is called. If a destination has not been reached, then at 870 the next hop node “k” is selected from neighbor nodes, message “m” is updated with the destination selection, and message “im” is sent to node “k”. At optional 875 the entry for the type “m” message may be updated.

[0067] The Q-learning next hop selection process at 870 is illustrated in more detail in FIG. 10. At 1010 an example selection criterion is set as a neighbor with minimum estimated cost, determined as k=minarg(C+Q(i)), where k is the neighbor with minimum estimated cost, C is local cost and Q(i) represents estimated cost for the total path from the neighbor node i. At 1020, the neighbor with the best estimation, namely k, is selected as the next hop with the probability (1−ε). The probability of selecting other neighbors as the next hop is ε.

[0068] In those cases in which the destination node sends a confirmation message back to the originating node, the confirmation message is processed according to the flow chart illustrated in FIG. 11. Confirmation management process 1100 begins with receipt of the MCBR confirmation message “m” from node “k” at 1110. At 1120 an entry is located for the type of the message and the entry for the message is updated. The Q-learning update process is accomplished through Q=αQnew+(1−α)Qold, where α is the learning rate and Q is the weighted sum of new cost values (Qnew) from the confirmation message and old cost values (Qold) in the entry.

[0069] In a Q-learning-based routing strategy, a backward phase is useful for updating the heuristics from neighbors. However, it only updates the value for one node, along the backward link. Optionally, a determination may be made at 1130 as to whether the current node is a source node or if there is no change in the message entry. If either of the above is true, there is no further back propagation. Otherwise, the confirmation message “m” may be sent back to its previous forwarding node 1140. Note that in this case, the forwarding routine should save the link to its previous node in the entry. Another option would be to make the backward phase update heuristic values along the path n-links back to the source, simultaneously with passing messages forward. It would be discontinued when there are no further changes. In this case, every entry saves a backward link, pointing to the node forwarding the message. This is described more fully in FIG. 12, where the back propagation process begins with receipt of the MCBR confirmation message “m” having a number (N) of back propagations. A determination is made at 1220 as to whether this is a new type message. If it is a new type message, at 1230 an entry is created for the message and the entry is updated with the Q value of the created entry. If the message is not a new type message, then at 1240 its existing entry is located, the entry is updated by updating the Q value of that entry as was done in FIG. 11 and the number of back propagations is reduced by 1, from N to N−1. At 1250 a decision is made whether to stop back propagation. The back propagation stops when (1) it is a source node or (2) Q value has no change at this update or (3) N=0, i.e., no more steps are left.

[0070] An example of extended backward propagation is illustrated in the pseudo code below, provided for illustrative purposes. It will be appreciated that the various embodiments of the invention disclosed herein could assume numerous expressions in such code, all of which are contemplated by the scope of the specification and claims herein.

received_Q (Qm) at node v from node w do
if (Qm(v, w) ≠ Q (m)) do
Qm(v, w)  Qm(v, w) + α( Qm − Qm(v, w))
send_Q (vm(v ) + min Qm(v, w)) to bm(v);
end
end

[0071] In the above pseudo code, bm(v) indicates the backward node of v for message m, i.e., v receives m from bm(v), which is recorded in the entry during the forward phase.

[0072] Variations of back propagation may be applied; for example, back propagating only when a message arrives at its destination, or back propagating for specified links or a specified number of links only. Adding backward propagation does not change the message complexity of delivery or route convergence; however, it may reduce time complexity for a single message delivery, or reduce the number of messages needed to converge, since heuristics are updated more quickly. Without back propagation, a message beginning at a node S may loop many times before finding the route to a destination node D.

[0073] While the present invention has been illustrated and described with reference to specific embodiments, further modification and improvements will occur to those skilled in the art. For example, the learning rate may be varied, which may improve the stability of the routing strategies while the network is dynamic. Additionally, “code” as used herein, or “program” as used herein, is any plurality of binary values or any executable, interpreted or compiled code which can be used by a computer or execution device to perform a task. This code or program can be written in any one of several known computer languages. A “computer”, as used herein, can mean any device which stores, processes, routes, manipulates, or performs like operation on data. It is to be understood, therefore, that this invention is not limited to the particular forms illustrated and that it is intended in the appended claims to embrace all alternatives, modifications, and variations which do not depart from the spirit and scope of this invention.

Claims

1. A method for a learning-based strategy utilized within message-initiated constraint-based routing for digital message communication among nodes in an ad-hoc network, wherein each node includes a plurality of attributes having attribute values, comprising: determining local attributes for each of the nodes in the ad-hoc network; providing access to the attributes of each neighboring node, wherein said neighboring node is a node one hop away from a current node within the ad-hoc network; estimating the minimum and maximum values of at least one attribute within the plurality of attributes; defining constraints on the plurality of attributes; performing constraint checking for each message type, wherein a message type includes a destination specification, route specification, and objective specification; performing cost estimation checking for each said message type, wherein cost is defined as the total value of an attribute along a message path; utilizing cost estimation to converge on an optimal message path; and routing a message within the ad-hoc network, said message including the address of a sending node, the address of the source node, route and destination constraints, the number of route constraints, the number of destination constraints, message identification number, sequence identification number, and routing objectives.

2. The method for a learning-based strategy utilized within message-initiated constraint-based routing according to claim 1, wherein said message further comprises a routing strategy identification number for identifying a routing strategy type.

3. The method for a learning-based strategy utilized within message-initiated constraint-based routing according to claim 2, wherein said routing strategy type comprises a search-based routing strategy.

4. The method for a learning-based strategy utilized within of message-initiated constraint-based routing according to claim 3, wherein said search-based routing strategy comprises: receiving not less than one forward message from not less than one source node within the ad-hoc network; determining whether said forward message is of a new type; identifying an appropriate entry for said forward message type; estimating a cost for transmission of said forward message from a receiving node to a destination node; sending a confirmation of said estimated cost; reviewing the destination of said forward message and processing said forward message if said destination has been reached; and reviewing the destination of said forward message and selecting a next node for receipt of said forward message, updating said forward message with identification of the intermediate node through which said forward message has passed, and transmitting said forward message to said next node if said destination has not been reached.

5. The method for a learning-based strategy utilized within message-initiated constraint-based routing according to claim 4, wherein estimating a cost for transmission from said receiving node to said destination node comprises determining the sum of the local cost and the estimated cost for the total path from said next node.

6. The method for a learning-based strategy utilized within message-initiated constraint-based routing according to claim 4, wherein sending a confirmation of said estimated cost comprises sending said estimated cost to said source node.

7. The method for a learning-based strategy utilized within message-initiated constraint-based routing according to claim 4, wherein sending a confirmation of said estimated cost comprises broadcasting said estimated cost to a plurality of nodes within the network.

8. The method for a learning-based strategy utilized within message-initiated constraint-based routing according to claim 4, further comprising sending a destination confirmation to said source node.

9. The method for a learning-based strategy utilized within message-initiated constraint-based routing according to claim 4, further comprising updating said entry for said message type.

10. The method for a learning-based strategy utilized within message-initiated constraint-based routing according to claim 4, wherein next node selection is based on a selection criterion, said selection criterion comprising the neighbor with the minimum estimated cost.

11. The method for a learning-based strategy utilized within message-initiated constraint-based routing according to claim 10, wherein said next node selected comprises a neighbor node having the highest probability of meeting said selection criterion.

12. The method for a learning-based strategy utilized within of message-initiated constraint-based routing according to claim 10, wherein the next node selected is a plurality of nodes within the network.

13. The method for a learning-based strategy utilized within message-initiated constraint-based routing according to claim 8, wherein sending a destination confirmation comprises: receiving a message confirmation from a node on the network; locating a message entry for the message type of said received message; and updating said message entry with the cost objective estimate, wherein said cost objective estimate comprises the weighted sum of the new cost values from said confirmation message and the previous cost values in the message entry.

14. The method for a learning-based strategy utilized within message-initiated constraint-based routing according to claim 8, further comprising: determining whether said receiving node is a source node or if there is no change in the message entry; and sending a confirmation message to said source node if either the current node is not a source node or if there is a change in said message entry.

15. The method for a learning-based strategy utilized within message-initiated constraint-based routing according to claim 8, further comprising updating cost values at all nodes in said message path back to said source node.

16. The method for a learning-based strategy utilized within message-initiated constraint-based routing according to claim 4, further comprising back propagating a confirmation message.

17. The method for a learning-based strategy utilized within message-initiated constraint-based routing according to claim 16, wherein said back propagation comprises: receiving a confirmation message having a number of back propagations, wherein back propagations updates said message entry for all receiving and sending nodes along said message path; determining whether said message is of a new type; creating a message entry if said message is a new type message; updating said message entry with a new cost value; locating an existing message entry for a message which is not a new type message; updating the cost value of said existing message entry and reducing the back propagation number count by one; discontinuing back propagation if the receiving node is a source node; discontinuing back propagation if said new cost value is the same as said updated cost value; and discontinuing back propagation if no additional potential receiving nodes are identified.

18. The method for a learning-based strategy utilized within message-initiated constraint-based routing according to claim 16, wherein back propagation is performed only when said message arrives at said destination node.

19. The method for a learning-based strategy utilized within message-initiated constraint-based routing according to claim 16, wherein back propagation is performed only for specified nodes along said message path.

20. The method for a learning-based strategy utilized within message-initiated constraint-based routing according to claim 16, wherein back propagation is performed only for a specified number of nodes along said message path.

21. A system for a learning-based strategy utilized within message-initiated constraint-based routing for digital message communication among nodes in an ad-hoc network, wherein each node includes a plurality of attributes having attribute values, comprising: local attribute module means for determining local attributes for each of the nodes in the ad-hoc network; remote attribute module means for providing access to the attributes of each neighboring node, wherein said neighboring node is a node one hop away from a current node within the ad-hoc network; timer module for providing a time trigger function for said local attribute module and said remote attribute module; broadcast module for providing a send function for said local attribute module and a receive function for said remote attribute module; attribute property module for estimating the minimum and maximum values of at least one attribute within the plurality of attributes; constraint module for defining constraints on the plurality of attributes, performing constraint checking for each message type, wherein a message type includes a destination specification, route specification, and objective specification, and performing cost estimation checking for each said message type, wherein cost is defined as the total value of at least one attribute along a message path; and routing module for routing a message within the ad-hoc network utilizing a learning-based strategy to converge on an optimal message path, said message including the address of a sending node, the address of the source node, route constraints, destination constraints, the number of route constraints, the number of destination constraints, message identification number, sequence identification number, and routing objectives.

22. The system for a learning-based strategy utilized within message-initiated constraint-based routing according to claim 21, wherein said learning-based strategy comprises cost estimation.

23. The system for a learning-based strategy utilized within message-initiated constraint-based routing according to claim 21, wherein said message further comprises a routing strategy identification number for identifying a routing strategy type.

24. The system for a learning-based strategy utilized within message-initiated constraint-based routing according to claim 23, wherein said routing strategy type comprises a search-based routing strategy.

25. The system for a learning-based strategy utilized within message-initiated constraint-based routing according to claim 24, wherein said search-based routing strategy comprises: means for receiving not less than one forward message from not less than one source node within the ad-hoc network; means for determining whether said forward message is of a new type; means for identifying an appropriate entry for said forward message type; means for estimating a cost for transmission of said forward message from a receiving node to a destination node; means for sending a confirmation of said estimated cost; means for reviewing the destination of said forward message and processing said forward message if said destination has been reached; and means for reviewing the destination of said forward message and selecting a next node for receipt of said forward message, updating said forward message with identification of the intermediate node through which said forward message has passed, and transmitting said forward message to said next node if said destination has not been reached.

26. The system for a learning-based strategy utilized within message-initiated constraint-based routing according to claim 25, wherein estimating a cost for transmission from said receiving node to said destination node comprises determining the sum of the local cost and the estimated cost for the total path from said next node.

27. The system for a learning-based strategy utilized within message-initiated constraint-based routing according to claim 25, wherein sending a confirmation of said estimated cost comprises sending said estimated cost to said source node.

28. The system for a learning-based strategy utilized within message-initiated constraint-based routing according to claim 25, wherein sending a confirmation of said estimated cost comprises broadcasting said estimated cost to a plurality of nodes within the network.

29. The system for a learning-based strategy utilized within message-initiated constraint-based routing according to claim 25, further comprising sending a destination confirmation to said source node.

30. The system for a learning-based strategy utilized within message-initiated constraint-based routing according to claim 25, further comprising updating said entry for said message type.

31. The system for a learning-based strategy utilized within message-initiated constraint-based routing according to claim 25, wherein next node selection is based on a selection criterion, said selection criterion comprising the neighbor with the minimum estimated cost.

32. The system for a learning-based strategy utilized within message-initiated constraint-based routing according to claim 31, wherein said next node selected comprises a neighbor node having the best possibility of meeting said selection criterion.

33. The system for a learning-based strategy utilized within of message-initiated constraint-based routing according to claim 31, wherein the next node selected comprises a plurality of nodes within the network.

34. The system for a learning-based strategy utilized within message-initiated constraint-based routing according to claim 29, wherein means for sending a destination confirmation comprises: means for receiving a message confirmation from a node on the network; means for locating a message entry for the message type of said received message; and means for updating said message entry with the cost objective estimate, wherein said cost objective estimate comprises the weighted sum of the new cost values from said confirmation message and the previous cost values in the message entry.

35. The system for a learning-based strategy utilized within message-initiated constraint-based routing according to claim 34, further comprising: means for determining whether said receiving node is a source node or if there is no change in the message entry; and means for sending a confirmation message to said source node if either the current node is not a source node or if there is a change in said message entry.

36. The system for a learning-based strategy utilized within message-initiated constraint-based routing according to claim 29, further comprising back propagating a confirmation message.

37. The system for a learning-based strategy utilized within message-initiated constraint-based routing according to claim 36, wherein said back propagation comprises: means for receiving a confirmation message having a number of back propagations, wherein back propagations updates said message entry for all receiving and sending nodes along said message path; means for determining whether said message is of a new type; means for creating a message entry if said message is a new type message; means for updating said message entry with a new cost value; means for locating an existing message entry for a message which is not a new type message; means for updating the cost value of said existing message entry and reducing the back propagation number count by one; means for discontinuing back propagation if the receiving node is a source node; means for discontinuing back propagation if said new cost value is the same as said updated cost value; and means for discontinuing back propagation if no additional potential receiving nodes are identified.

38. The system for a learning-based strategy utilized within message-initiated constraint-based routing according to claim 36, wherein back propagation is performed only when said message arrives at said destination node.

39. The system for a learning-based strategy utilized within message-initiated constraint-based routing according to claim 36, wherein back propagation is performed only for specified nodes along said message path.

40. The system for a learning-based strategy utilized within message-initiated constraint-based routing according to claim 36, wherein back propagation is performed only for a specified number of nodes along said message path.

41. An article of manufacture comprising a computer usable medium having computer readable program code embodied in said medium which, when said program code is executed by said computer causes said computer to perform method steps for a learning-based strategy utilized within message-initiated constraint-based routing for digital message communication among nodes in an ad-hoc network, wherein each node includes a plurality of attributes having attribute values, the method comprising: determining local attributes for each of the nodes in the ad-hoc network; providing access to the attributes of each neighboring node, wherein said neighboring node is a node one hop away from a current node within the ad-hoc network; estimating the minimum and maximum values of at least one attribute within the plurality of attributes; defining constraints on the plurality of attributes; performing constraint checking for each message type, wherein a message type includes a destination specification, route specification, and objective specification; performing cost estimation checking for each said message type, wherein cost is defined as the total value of an attribute along a message path; utilizing cost estimation to converge on an optimal message path; and routing a message within the ad-hoc network, said message including the address of a sending node, the address of the source node, route and destination constraints, the number of route constraints, the number of destination constraints, message identification number, sequence identification number, and routing objectives.