Researchers have developed Faynt, a family of Transformer-based AI policies designed for the competitive video game Super Smash Bros. Melee. These policies, with 10 million and 75 million parameters, can control all 26 characters using a single model checkpoint. The 10M parameter model demonstrates a high win rate against various specialist and multi-character AI opponents, and even a zero-delay AI model. The development involved extensive pretraining on human replays, followed by optimization techniques like distillation and reinforcement learning. AI
IMPACT Sets a new benchmark for AI performance in complex real-time strategy games, potentially influencing future research in game AI and reinforcement learning.
RANK_REASON The cluster describes a new AI model and its performance on a specific task, published on arXiv. [lever_c_demoted from research: ic=1 ai=1.0]
- arXiv
- Faynt
- Hugging Face
- NVIDIA T4
- reinforcement learning
- Slippi-AI
- Super Smash Bros. Melee
- Transformer
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
