Training through self-play
We train policies through repeated interaction and competition, using feedback to improve decisions over successive rounds of training.
Reinforcement learning · Adaptive agents
Intelligence, learned
through experience.
We build learning systems that turn interaction into capability. Our work combines reinforcement learning, memory, and self-play to train agents that improve through experience.
From training to competition
First place at generals.bot
We trained and deployed a self-play agent that took first place in the generals.bot competition. The system combines recurrent reinforcement learning, a custom self-play pipeline, and an optimized CPU inference implementation. We are now optimizing our training system.
Decisions with memory
Our recurrent agent uses memory to make decisions from partial observations. This is a starting point for studying behavior over longer horizons.
Potential applications
Our current focus is improving the system's training and performance. The methods may later find applications in language agents and robotics.