Brandubh

Rules:

- All pieces (king included) move like rooks in chess, horizontally or vertically as far as they want.
- The attackers (black) win by capturing the white king (the center piece).
- The defenders (white) win by escaping with the king to one of the four corners.
- Both sides (king included) captures enemy pieces by, through a move, squizing them in between two of your pieces (horizontally or vertically).
- The four corners can only be occupied by the king, and can also be used as an allied piece in a capture.
- Three-fold repetition does not result in a draw, and is instead an illegal move.

AIs:

- The "NN + MCTS" AI is inspired by AlphaZero, and is trained with reinforcement learning, and contains a neural network (NN) which predicts good moves and the game evaluation. It provides the most natural and human-like play. It performs a monte carlo tree search (MCTS), guided by its NN, to look for good moves. The difficult setting decides the number of simulations: Easy (100), Medium (200), Hard (400), Very Hard (800).
- The "Minimax" AI is inspired by traditional engines like Stockfish, and simulates all possible game states for X moves into the future. It relies of a set of hard-coded rules for good play (called "heuristics"). These are not tuned well, and this AI therefore plays quite strangely until it finds a forced winning move. The difficulty decides what depth into the future it will look: Easy (3), Medium (4), Hard (5), Very hard (6).

Neural Network Settings

ONNX Model:

Select which neural network checkpoint to use. The model will reload automatically.

Temperature: 0.4

Randomness in AI move selection. 0 = deterministic (always picks best move). Set to 0.4 - 0.6 for dynamic yet strong play.

MCTS Simulations (Eval/Suggestions): 200

Number of simulations for evaluation bar and move suggestions. Higher = more accurate but slower. Does not affect AI player difficulty.

FPU Reduction: 0.50

First Play Urgency reduction (relative to parent). Controls how optimistic/pessimistic MCTS is about unvisited nodes. Lower values (0-0.3) = more optimistic exploration. Higher values (0.7-1.5) = more cautious. Default 0.5 balances exploration and exploitation.

C_puct (Exploration): 1.20

Exploration constant for PUCT formula. Controls exploration vs exploitation trade-off. Lower values (0.5-1.0) = more exploitation of known good moves. Higher values (1.5-3.0) = more exploration of uncertain moves. Default 1.2 provides balanced search.

Continuous Eval/Suggestions Update

Continuously run MCTS and update eval bar and move suggestions every 100 simulations. Ignores the simulation count parameter and keeps improving until position changes.