Agent仿真中具有先验知识的混合学习算法与混合结构模型

A hybrid learning algorithm with prior knowledge and a hybrid architecture model in agent-based simulation

强化学习是一种有效的机器学习方法,是无监督学习,通过不断地和环境交互得到外部环境评价信号,选择合适的动作.Q学习是一种典型的强化学习,其学习效率较低,尤其是当状态空间和决策空间较大时.为提高Q学习学习效率和收敛速度,采用具有先验知识的Q学习算法,利用模糊综合决策方法处理专家经验和环境信息得到Q学习的先验知识,对Q学习的初始状态进行优化;针对Agent个体学习与群体学习各自的不足,提出了采用混合学习算法,将个体学习与群体学习有效结合起来,提高了Agent的个体性能及系统整体的智能水平;同时为满足复杂适应性需求,采用Agent混合结构模型,在该模型中构造了基于知识的协调控制器,通过它来协调慎思式过程和反应式过程.

Reinforcement Learning is an effective learning method of Machine learning, it has no supervision, and it can choose optimal actions by continuously interacting with environment. Q-learning is a typical Reinforcement Learning (RL) method with a slow convergence speed especially as the scales of the state space and action space increase. An improved Q-learning method using prior knowledge uses fuzzy integrated decision-making to process expert knowledge, which optimizes the initial states to give better learning foundation. A hybrid learning algorithm based on improved Q-learning is proposed to combine individual learning and group learning effectively. It improves agent’s ability and system’s intelligence level. To satisfy the requirements of complex adaptability, a hybrid architecture model is developed. In this model, a coordination control unit based on knowledge is proposed to coordinate the cognitive process and reactive process.