Modifying MCTS for human-like general video game playing

Abstract

We address the problem of making general video
\ngame playing agents play in a human-like manner.
\nTo this end, we introduce several modifications of
\nthe UCT formula used in Monte Carlo Tree Search
\nthat biases action selection towards repeating the
\ncurrent action, making pauses, and limiting rapid
\nswitching between actions. Playtraces of human
\nplayers are used to model their propensity for repeated
\nactions; this model is used for biasing the
\nUCT formula. Experiments show that our modified
\nMCTS agent, called BoT, plays quantitatively similar
\nto human players as measured by the distribution
\nof repeated actions. A survey of human observers
\nreveals that the agent exhibits human-like playing
\nstyle in some games but not others.