Faynt transformer policies win many games in Super Smash Bros Melee
Faynt: Scaling and Optimizing Policies for Competitive Melee
Machine Learning
Summary
Playing the video game Super Smash Bros. Melee well is hard because it has many characters and fast gameplay. The authors created a computer program called Faynt using transformer models that can control all 26 characters in the game. After training with human gameplay data and playing against other strong game-playing programs, Faynt wins most matches, showing strong and consistent play. The program can make quick decisions and is available for others to try and improve through open-source tools.
What this means in practice
- •For game ai developers: Use Faynt’s multi-character transformer policy design and training methods to build competitive agents for complex fighting games with diverse characters.
- •For video game platform engineers: Incorporate efficient model inference techniques from Faynt to enable real-time decisions in high-speed game AI environments.
Authors
Ali Janati, Nikita Kuzmin, Rohit Swamy, Charles Niu
Abstract
We introduce Faynt, a family of 10M- and 75M-parameter Transformer policies for Super Smash Bros. Melee, each controlling all 26 characters with a single checkpoint. After reinforcement learning (RL), the 10M wins 240 of 244 same-character games (98.4%) against fourteen specialist and multi-character releases on their supported rosters, with a winning record against every release. These opponents retain 21- or 24-frame action delays; Faynt uses no added delay, and we have not isolated the effect of this difference. In a separate evaluation against a privately supplied zero-delay Slippi-AI model, the 10M wins all 68 games across two conditioning settings. We study architecture, optimization, scaling, and hyperparameter transfer to guide pretraining on approximately 840,000 human replays. Post-training combines rank- and outcome-based curricula, 75M-to-10M distillation, and RL restricted to Fox mirror matches. On the initial 152-game benchmark, the supervised 10M wins 69.7% of games, compared with 45.4% for the pretrained 75M, despite higher overall held-out controller-prediction loss. The weighted validation loss used for supervised checkpoint selection agrees with the win-rate ordering of all four pretrained and supervised policies. After supervised post-training, both models take less damage per minute, build larger early leads, and win more often after losing the first life. Optimized inference on recorded game states averages 5.2 ms per decision for the 10M and 8.7 ms for the 75M on an NVIDIA T4, excluding emulator execution and communication. We open-source the weights, both benchmark suites, and a platform for automated model tournaments.