<ul><li>Assistance games are a promising alternative to reinforcement learning from human feedback for training AI assistants.</li><li>A new approach called AssistanceZero is presented, which extends AlphaZero with a neural network to solve assistance games in complex environments.</li><li>AssistanceZero outperforms model-free RL algorithms and imitation learning in a Minecraft-based assistance game.</li><li>In a human study, the AssistanceZero-trained assistant significantly reduces the number of actions required to complete building tasks.</li></ul>

AssistanceZero: Scalably Solving Assistance Games

Discover more