27 Nov 2015 | Ardi Tampuu*, Tambet Matiisen*, Dorian Kodelja, Ilya Kuzovkin, Kristjan Korjus, Juhan Aru†, Jaan Aru, Raul Vicente‡
This paper explores the emergence of competitive and collaborative behaviors in multiagent systems using Deep Q-Learning Networks (DQNs) in the classic video game Pong. The authors manipulate the rewarding scheme of Pong to demonstrate how agents controlled by independent DQNs interact and evolve their strategies. In the fully competitive mode, agents learn to score efficiently, while in the fully cooperative mode, agents focus on keeping the ball in play for as long as possible. The study also investigates the transition between these modes by adjusting the reward values for scoring and losing the ball. The results show that DQNs can effectively learn decentralized strategies in complex environments, making them a practical tool for studying multiagent systems. The paper discusses the limitations of the current approach, such as overestimation of Q-values in competitive scenarios, and outlines future directions, including the study of larger agent populations and the emergence of communication codes in complex environments.This paper explores the emergence of competitive and collaborative behaviors in multiagent systems using Deep Q-Learning Networks (DQNs) in the classic video game Pong. The authors manipulate the rewarding scheme of Pong to demonstrate how agents controlled by independent DQNs interact and evolve their strategies. In the fully competitive mode, agents learn to score efficiently, while in the fully cooperative mode, agents focus on keeping the ball in play for as long as possible. The study also investigates the transition between these modes by adjusting the reward values for scoring and losing the ball. The results show that DQNs can effectively learn decentralized strategies in complex environments, making them a practical tool for studying multiagent systems. The paper discusses the limitations of the current approach, such as overestimation of Q-values in competitive scenarios, and outlines future directions, including the study of larger agent populations and the emergence of communication codes in complex environments.