基于BDI框架的多Agent动态协作模型与应用研究

近年来，多Agent学习已经成为人工智能和机器学习研究方向发展最迅速的领域之一．将强化学习和BDI思维状态模型相结合，形成针对多Agent的动态协作模型．在此模型中，个体最优化概念失去其意义，因为每个Agent的回报，不仅取决于自身，而且取决于其它Agent的选择．模型采用AFS神经网络对输入状态空间进行压缩，提高强化学习的收敛速度．与此同时，利用模拟退火算法启发性地指明动作空间搜索方向，使其跳出局部最小点，避免迭代步数的无限增长．理论分析和在机器人足球领域的成功应用，都证明了基于BDI框架的多Agent动态协作模型的有效性。

A MULTIAGENT DYNAMIC COOPERATING MODEL BASED ON BDI FRAMEWORK ANDITS APPLICATION

Multiagent learning has become one of the rapidest development fields in AI and machine learning in recent years. In this paper, a multiagent dynamic cooperating model is proposed, which connects reinforcement learning with the BDI model. Notions of individual optimality loses its meaning since each agent's payoff depends not only on iteself but also on other agent's choices. The model adopts AFS NN to compress input state space, which can improve the reinforcement learning's convergence velocity. At the same time, the simulated annealing algorithm is applied to heuristically point the search direction of action space. It can leap out the local minimum point and avoid the infinite growth of the iteration step number. Theory analysis and success in the robot soccer domain have proved the effectiveness of the multiagent dynamic cooperating model based on the BDI framework.