Collaborative strategies of partially interacting agents in a partially observable universe

We are interested in optimizing the strategies of autonomous agents, which are coordinating their actions toward a same objective. Partial observations of the universe and partial communications between the agents are hypothesized. Thus, the strategies of the agents are partially independent. This problem is related to the POMDP family. However, our approach is not based on the dynamic programming. In a previous work, it has been shown that the optimal strategies for the POMDP problem could be approximated by dynamic Bayesian networks with input and output (i.e. exchanging information with the universe). A method, based on the cross-entropy, has been implemented for optimizing the parameters of such DBN, relatively to the POMDP problem; as a result the whole method was able to compute a near-optimal solution of the POMDP. This method is applied to our problem of autonomous agents. In this case, the strategy of each agent is modelled by a dynamic Bayesian network. These DBNs are assumed to be disjoint (independence of the agents). However, these DBNs are able to exchange information with the universe and with the communication system.