Description

The goal is a reinforcement learning game AI that plays the official Pokémon battle format (VGC) at a competent level. Because the sample space of possible moves and abilities is vast compared to games like chess, we use Monte Carlo Tree Search (MCTS) to guide moves rather than a pure reinforcement learning approach that would have to learn the rules of the game on its own. The result is similar in spirit to what Lc0 is for chess.

Our objective is to complete the VGC RL agent from last semester, reaching performance parity with the VGC Bench paper, and then deploy it so that anyone can click a link and play a game against the bot.

That involves:

• Implementing the Phasic Policy Gradient algorithm
• Writing a distributed training loop with a shared policy, trained on the cluster
• Writing a team preview bot to decide team selection
• Hyperparameter optimization
• Collecting gameplay data for behaviour cloning at the start of training
• Dockerizing the model and server, then deploying online
• Possibly collecting data from users who play the bot, to improve further

Prerequisites

• Python (NumPy, PyTorch, any plotting library)
• RL basics — actor-critic models, implemented in PyTorch
• Node.js basics (self-hosting websites)