Patent Application
20080040178
The manner in which the invention is implemented and its attendant advantages will become clear in the following description of the implementation method, reference being made to the appended figures, in which:
a to 5d are representations of a tree forming part of a collaborative network between agents, which show the various stages of the resolution process.
As we have already seen, the invention concerns a process whereby negotiations between different agents needing to share critical resources can be organized. This process applies to numerous applications, and the example described below, given purely for illustrative purposes, in no way limits the scope of the invention.
By the same token, the production planning example in question has been highly simplified to make the invention easier to understand, but needless to say, the benefits of the latter naturally reside in the solving of much more complex problems.
In the case illustrated by the industrial process scheduling figures, the various tasks to be accomplished are defined as “agents” needing to share common “resources”, namely the uptime of a machine on which the tasks need to be performed. The two top timing diagrams in
In the initial state, before the invention's process is applied, tasks T1 and T2 are scheduled with a certain time lag, and the production goal is to allow tasks T2 and T3 to be performed on machines M1 and M2.
Among the additional constraints or aims of the example is the fact that task T4 can only be executed on machine M2, whereas task T3 may be executed equally on either of the machines M1 or M2.
In accordance with the invention, the first stage consists in setting up a collaborative network between the various agents liable to share a common resource. In this instance, tasks T1-T4 are likely to share the common uptime resource of machines M1 and M2. The collaborative network is formed by listing the possibilities of exchanging a given resource between tasks, in other words switching the uptimes of machine M1 or M2. In this way, the network expresses all the relations between agents, and constitutes an infrastructure whereby the various agents will express their satisfaction rating in respect of each switchover during the negotiations.
Naturally, agents may be linked by other relations to other agents, who are not shown inasmuch as they play no part in resolving the particular problem of planning execution of the four tasks T1-T4. Furthermore, for this type of problem, one could show certain time constraints between the tasks to be assigned to the machines.
As shown in
In this resolution process, an iterative succession of stages is performed, in which each agent perceives his satisfaction rating S2, according to the validity verdicts on the various relations he is part of, and which make up the leaves of the tree of which he is the root. More precisely, the various basic stages unfold in a cycle shown in
At each node J10, J11, J12 of this tree, a calculation is performed according to the downstream nodes and/or extremities, involving a mass of variables, such as the detection of minimums, maximums or mean values. In the instance shown in
Subsequently, each agent performs an action on the tree of which he is the origin, by selecting certain contracts that can satisfy him, marking his environment to encourage reciprocal selection of the contracts, in other words by issuing a valid verdict, according to stage S4 of
So, as
Furthermore, as shown in
At each cycle, each agent thus perceives the validation verdicts for the relations matching the nodes of potential collaboration with other agents and, as explained earlier, marks his environment according to the selections he prefers.
This sequence of stages S2, S3, S4 continues until overall resolution is achieved according to test S5.
As shown in
The resolution as described above reinforces links having high satisfaction ratings. However, in a manner not illustrated herein, a fraction of near-random behaviour can be instilled into the resolution process with a view to encouraging the emergence of solutions that are in theory incompatible with a would-be satisfactory solution, but only locally; in which case, satisfaction ratings computed using a determinist law (by calculating a maximum or minimum value, for instance) may be upset by a random law. This could for instance consist in randomly selecting one of the sub-trees that would not have been selected by the determinist law. This approach favours an exploratory process aimed at averting the deadlocks arising when satisfactory solutions appear only locally rather than globally.
As we have demonstrated, the invention presents numerous advantages, inasmuch as it allows complex exchanges between multiple agents through negotiations supported by the environment, commonly termed stigmergic negotiations; indeed, rather than negotiating directly with one another, the agents negotiate by marking their local environment, in other words by issuing satisfaction ratings and validity verdicts. This type of stigmergic communication can be likened to that of social insects, which communicate by marking their environment, for instance through chemical structures such as pheromones. This process proves to be advantageous insofar as it can be adapted to many situations and applied to different types of systems to determine a satisfactory solution (or solutions if they exist) from a global standpoint.
The invention can be used to work out complex solutions to difficult problems, problems that are still hard to resolve through other MAS approaches. Thanks to this invention, the robust, flexible and decentralized features specific to multiagent systems can therefore be used for critical resource sharing problems.
One advantage of the invention is that it proposes a generic approach. The collaborative network can be used to represent a wide variety of critical resource sharing problems, which can be resolved on the basis of the same agent behaviours thanks to the system's capacity to organize itself.
