Competition and collaboration among fuzzy reinforcement learning agents

GARIC-Q, introduced earlier by Berenji (1996), performs incremental dynamic programming using intelligent agents which are controlled at the top level by fuzzy Q-learning and at the local level, each agent learns and operates based on GARIC, a technique for fuzzy reinforcement learning. In MULTI-GARIC-Q, agents can collaborate to investigate the state space by training a unique action evaluator which acts as a critic function for all the agents. MULTI-GARIC-Q2 extends MULTI-GARIC-Q by providing more independence to each fuzzy reinforcement learning agent to investigate its state space for an extended time. After the interim learning period is over, all the agents engage in a competition to select a winner who will then continue its learning using the actual dynamic system. The rest of the agents continue learning using copies of a simulated plant. MULTI-GARIC-Q2 provides faster learning while allowing each agent to act and learn more independently.