Q-learning with prior knowledge in multi-agent systems

Abstract

Reinforcement Learning (RL) is an important branch of machine learning and it is unsupervised without specific signals. The learning process adjusts its actions according to external signals from interactions with the environment as a result, the system learning speed is relatively slow. Q-learning is a typical RL method with a slow convergence speed especially as the scales of the state space and the 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 a better learning foundation. Test results on the Robot Soccer system show that the improved Q learning method has a higher learning efficiency and convergence speed.