Chinanews.com, Beijing, December 7th (Reporter Sun Zifa) The latest artificial intelligence (AI) paper published in the academic journal "Nature-Communications" under Springer Nature reports that artificial intelligence agents can negotiate and form agreements so that It outperformed other incapable agents in the board game Diplomacy.
The research demonstrates a deep reinforcement learning approach that models agents that can communicate and cooperate with other artificial agents to jointly plan while playing a game.
According to the paper, it is important to develop artificial intelligence that can exhibit cooperation and communication between agents.
Diplomacy, a popular board game that provides a useful testbed for such behavior, involves complex communication, negotiation and alliance formation between players that has been difficult for artificial intelligence to achieve.
To win the game, Diplomacy needs to deduce the future plans of the players present, the commitments between players, and their honest cooperation.
Past AI agents have had success in single-player or two-player games where there is no communication between the players.
Lead author János Kramár, corresponding author Yoram Bachrach and colleagues from the UK-based artificial intelligence company DeepMind devised a deep reinforcement learning method that allows agents to negotiate alliances , to make plans together.
They created agents that simulate players, forming teams and trying to outwit other teams' strategies.
By predicting likely future game states, this learning algorithm allows agents to agree on future actions and identify beneficial trades.
To approach human-level performance, the authors also investigate the conditions for honest cooperation by examining some inter-agent commitment-breaking scenarios (where agents deviate from past agreements).
The authors conclude that their findings help AI agents form the basis for flexible communication mechanisms that allow them to adapt their strategies to the environment.
Furthermore, these findings also show how a propensity to sanction peers for violating agreements greatly reduces the advantage of such violators, contributing to mostly believable exchanges even though conditions initially favor the violators.
(Finish)