A variety of search-based implementations of an AI agent for the board game mancala (kalah).
The following files are taken from course content provided by Tim French from the University of Western Australia:
- Mancala.java - Game harness
- MancalaAgent.java - Interface for the AI agent API
- RandomAgent.java - An implementation of a simple random choice agent
All other files are implementations of the AI agent API by Jesse Wyatt. Listed in order of complexity:
- MMAgent.java - Simple minimax agent without pruning (8-ply fixed)
- ABAgent.java - Minimax search with alpha-beta pruning (10-ply fixed)
- ABWMAgent.java - Minimax search with alpha-beta pruning and transposition tables (12-ply fixed)
- ABIDAgent.java - Time limited iterative deepening extension of ABWMAgent (100ms soft-limited)
- MTDFAgent.java - MTD-f ("zero-width" iterative deepening) extension of ABWMAgent (100ms soft-limited)
Significant code is duplicated between agents because of practical academic restrictions and this may result in additional bugs in less complex agents. Code was initially written with OpenJDK 1.8 as a target, but later agents are modified to maintain compatibility with Java 1.5 and avoid the use of System library calls.
A game may be run by compiling all files and running java Mancala. Agent selection is hardcoded within Mancala.java with a default configuration of MTDFAgent vs RandomAgent. Other agents can be tested by modifying their types at instantiation in Mancala.main() before/during the call to Mancala.play(). Turn timers are limited to 0.5s and simpler agents with fixed search depths may potentially fail on less powerful machines.
Repository was shifted to public visibility as of 20/10/2018.