Distributed reinforcement learning for sequential decision making

The paper addresses a problem of reinforcement learning in a homogeneous non-communicating multi-agent system for sequential decision making. We introduce a particular reinforcement learning model composed of evidential reinforcement neural networks representing agents, a fusion center, and a decision maker. The fusion center combines beliefs in each hypothesis under consideration generated by the agents and produces pignistic probabilities of the hypotheses under consideration. These pignistic probabilities are used by a decision maker in a sequential pignistic probability ratio test to choose one of two actions: "defer decision" or "decide hypothesis k". The test is shaped to encourage early decisions and incorporates a finite decision deadline. Upon each decision, a non-binary reinforcement signal is computed by the environment, and is then fed back to the agents, which utilize it to learn an optimizing belief function. The learning algorithm adapts the "profit sharing strategy" to the sequential decision making setting.