(The British University in Dubai (BUiD), 2010-01) Al Hashimi, Rashid
Multi-agent reinforcement learning is a common method for optimizing agents' local decision in a distributed and scalable manner. However, the study and analysis of the state-of-the-art multi-agent reinforcement learning (MARL) algorithms have been limited to small problems involving few number of learning agents.The purpose of this project is to conduct an extensive evaluation and comparison of MARL algorithms when used in networks that exhibit the scale-free property. The Internet and the social network of collaboration in science are only few examples of real-world networks that exhibit this property. Toward this goal, we developed a simulator that facilitates studying combinations of MARL algorithms, strategic games and networks with control propagation via tokens. These tokens are considered an opportunity for agents to play. Tokens also initiate a factor of randomness in the environment given its probability distribution over agents. Preliminary experimental results showed a signi cant reaction to the increase of tokens when agents play battle of the sexes in Neural network; the increase in token transfer probability yields a higher reward and a faster conversion.