What this document is
Multiagent Cooperation and Competition with Deep Reinforcement Learning is a paper published in PLOS ONE on April 5, 2017, by Ardi Tampuu, Tambet Matiisen, Dorian Kodelja, Ilya Kuzovkin, Kristjan Korjus, Juhan Aru, Jaan Aru and Raul Vicente. The authors extend Deep Q-Learning, a reinforcement-learning method that lets a computer program learn to act from raw visual input alone, into settings with two learning agents instead of one, using the classic video game Pong as the test environment.
By changing how points are awarded in Pong, the authors show that the same underlying learning method can produce either competitive or cooperative behavior between the two agents, and that the balance shifts predictably as the incentive to cooperate increases. They also find that an agent trained against another adaptive, learning opponent plays more reliably across situations than one trained only against a fixed, hard-coded algorithm.
The paper is presented as a demonstration that Deep Q-Networks can serve as a general tool for studying how multiple independent learning systems behave when they share an environment, an approach with applications beyond games in any setting where autonomous software agents interact.
Where to find it
Available from PLOS ONE.