AI Agents - WolfgangKonen/GBG GitHub Wiki

Agents

Currently implemented agents

• Max-N agent
• Expectimax-N agent
• MC (Monte Carlo) agent
• MCTS (Monte Carlo Tree Search) agent
• MCTS Expectimax agent
• TD (Temporal Difference) agent
• TD n-tuple agent
• SARSA n-tuple agent
• Q-Leraning n-tuple agent
• HumanAgent
• RandomAgent

Additionally, some games have specific agents, mainly for test and evaluation purposes:

• Edax, HeurPlayer, BenchPlayer (for game Othello only)
• BoutonAgent (for game Nim only )
• AlphaBetaAgent (for game Connect-4 only)
• BasicStrategyBlackJackAgent (for game BlackJack only)
• DAVI2Agent, DAVI3Agent (for game RubiksCube only)

Note that Edax in GBG is currently only available on Windows-based systems, since it involves a special EXE file.

Wrappers

Each agent can be wrapped during {train, play, eval, compete, inspect} with

• an n-ply look-ahead tree search (either deterministic: Max-N, or non-deterministic: Expectimax-N)
• an MCTS-Wrapper (deterministic) or MCTSE-Wrapper (non-deterministic), inspired by AlphaZero

Activate Max-N-Wrapper by selecting Wrapper Mode MaxNWrapper in Tab Wrapper pars and by setting Wrapper nPly to a value larger than 0. But caution: larger values for Wrapper nPly may quickly lead to very long execution times, depending on the branching factor of your game.

Activate MCTS-Wrapper by selecting Wrapper Mode MCTSNWrapper in Tab Wrapper pars and by setting suitable values for Wrapper MCTS (# iterations), PUCT for WrapM and Depth for WrapM.