A Multi-Agent Approach to Bluffing
Tshilidzi Marwala, Evan Hurwitz · 2009
While the exact nature of bluffing is still unknown, it has been shown that a system involving agents capable of learning adaptively not only from the game being played, but also from their opponents, is in fact able to learn to predict its opponent's reactions. This knowledge in turn changes the statistical nature of a game being played, allowing agents to learn to bluff, based purely on rational reasoning, lending strong support to the theory that bluffing is simply playing the odds, and not an illogical, psychologically based action. The use of the Re-enforcement learning paradigm (Sutton & Barto, 1998), along with the TD() algorithm for adaptively training neural networks, has been shown to meet all of the requirements to produce such agents. Lastly, the design of the agent "view", has been seen to be the most important facet of creating bluffing agents, since their view of the game as inclusive of the other players allows for the incorporation of those players into its estimation of the game's outcome. With all of these steps adhered to, artificially intelligent agents can learn to bluff!