Deep RL Masters Asymmetric Baghchal Game Strategy

Baghchal, a traditional Nepali asymmetric board game, presents a complex strategic challenge for AI due to its distinct player roles: four tigers aiming to capture goats, and twenty goats attempting to immobilize the tigers. Despite its strategic depth and cultural significance, it has been underexplored in deep reinforcement learning (RL) literature. This study systematically investigates four prominent deep RL algorithms—Deep Q-Network (DQN), REINFORCE, Proximal Policy Optimization (PPO), and MuZero—training them on one side of the game and evaluating their performance on both. The algorithms were assessed based on win rate, draw rate, average captures, training convergence, and computational cost. MuZero emerged as the top performer, achieving an 86% win rate as the Tiger and 62% as the Goat, attributed to its model-based planning capabilities via Monte Carlo Tree Search. PPO also proved highly competitive across both asymmetric tasks, offering significantly reduced computational costs compared to MuZero. Analysis of emergent strategies revealed that model-based approaches excel in long-horizon planning, whereas value-based methods like DQN showed a bias towards the Tiger role due to stronger reward signals.