Developer trains a small AI on a single RTX 3080 Ti gaming GPU to 'play' Pokémon Red

Developer stmonty has trained a compact AI model on a single RTX 3080 Ti graphics card to play Pokémon Red. The project, detailed in a blog post titled "Teaching a World Model to Play Pokémon," involves training a world model based on LeWorldModel research, created by AI scientist Yann LeCun's new venture, AMI Labs. The small model, requiring only about 12.5 million parameters, was trained locally on a consumer-grade graphics card to teach the model the functions of each button press through screen observation.

Rather than predicting the entire next screen, the model uses compressed summaries with just 192 numbers. Initially, the model underwent "reward-free" training, learning without awareness of the game's winning conditions. During training, the model analyzed 42,382 grayscale frames over 1,000 short runs, utilizing a save file from Professor Oak's lab, including scripted routes, random button presses, and random wandering.

The model was trained from scratch, with 14 button presses planned and attempted in each run, and 512 plan samples per round, selecting the best 64. Despite initial challenges, the model successfully selected Squirtle as a starter Pokémon and achieved this over 100 runs, compared to zero with random button presses and only one with an untrained model.

Stmonty notes that while pressing A repeatedly works from the save, the model's ability to discover the solution independently is the key achievement. The next step is to guide the model from the lab to Professor Oak's lab, through dialogue, and finally to choose a Pokémon. While the developer believes that scaling up the model would not be enough to complete the game, the project's open-source code is available on GitHub as lePokeRed, and it is recommended to use a CUDA GPU for training.