Improvement of Learning of Agents with Communication Function in Pursuit Problem of Real-Number Environment

Abstract

Pursuit problem is one of the benchmark problems of multi-agents. For more realistic model, we have proposed a staged view and then a staged binocular view for hunters. Simulation results with these views show that hunters can even learn and capture the prey with a fuzzy Qlearning. In the previous study, we have proposed hunters with a communication function and compared the capture ability in several cases of sending information with a fuzzy Q-learning, where an action from Q-table for visual information and that for communication information are combined with the constant weights. The results were worse than that for only visual information. In this study, we dynamically change the weights of combination for a situation, where we learn the weight with a fuzzy Q-learning. The result gets better than that for only visual information.