Research On Multi-Agent Cooperative Problem Solving Methods

Multi-agent cooperation is an important research area of multi-agent systems (MAS). It is also an effective solution method in MAS. This paper summarizes the progresses and existing problems on researches on multi-agent cooperative problem solving, and presents results on the research of logic model of multi-agent cooperation, cooperative problem solving methods and evaluation of cooperative plans. Includes the following details:(1) Established a logic model of multi-agent cooperative problem solving.The logic model relates the mental state of BDI agents to the environment state,the language and its formal semantics combining with the environment are also introduced into this model; after representing the layered structure of agent's ability named as "able to do", "fit to do" and "can do", the whole cooperative problem solving process is redefined as five stages, including the representation of task allocation and decomposition, which has partially improved the work of Jennings and Panzarasa.(2) Gave the representation and evaluation of multi-agent cooperative plan based on influence diagram and situation calculus. In this part of work, multi-agent cooperation model has been established based on influence diagrams and the multi-agent cooperative plan has been represented as a program of ConGolog, a concurrent programming language based on the situation calculus. Upon this, the computational algorithm evaluating this joint strategy was given by the structural operational semantics of ConGolog. The reasonability on possibility distribution of nature reactions has also been discussed and it has shown that if the monotony of the assumed possibility distribution is the same as the actual one, different joint strategies can also been compared and evaluated correctly. It is an improvement to the works of Koller and Nillosn. (3) Proposed a multi-agent reinforcement learning method based on role tracking.Role property has been added in multi-agent learning model while extending single agent learning process to group agent learning process and a reinforcement learning algorithm based on role tracking has been given. Additionally, the reasonability and convergence of this algorithm has also been discussed and verified by experiment, which improves the work of Bowling and Littman.(4) Gave an efficient method to solve the problems of Factored Markov Decision Process model using feature vector extraction.Since the specialities of Factored Markov Decision Process (FMDP), some efficient algorithm has been given to address the problem of curse of dimensionality in large FMDP by approximating state value function through feature vector extraction. A key contribution of this part of work is that it reduces the computation complexity by reducing constraints in linear programming and speeds up learning rate and building rate of joint strategy by transplanting state value functions to the more complex game in reinforcement learning. Experiments on have also been provided to demonstrate the efficiency and correctness of this method which improves partially the work of Gestrin.