Multi-Agent Active Perception Based on Reinforcement Learning and POMDP

In this article, we address a form of active perception characterized by curiosity-driven, open-ended exploration with intrinsic motivation, carried out by a group of agents. The multiple agents and a large number of possible actions are the main motivation for incorporating Multi-Agent Reinforcement Learning used to train a neural network in order to derive agent's policy. Partially Observable Markov Decision Process framework is used to accommodate the inaccuracy of sensors and probabilistic nature of agent's actions. The proposed method incorporates a consensus that derives the common belief vector, thus allowing each agent to make its decisions based on information acquired by all agents involved in the process of active perception. A well-known benchmark problem with a decentralized tiger scenario was used to demonstrate the possibility of the method to generate agents with different perceptual characteristics by simply changing the agents' reward function related to their intrinsic motivation. The main validation of the proposed approach was performed by using an example of multi-agent search mission. Final results are presented and discussed, and possible avenues for progress on open problems are identified.