Abstract

Inverse reinforcement learning (IRL) is a technique for inferring human preferences from demonstrations of a target behaviour. Classical approaches make strong assumptions on human rationality, are designed for only a single agent and do not scale to high-dimensional environments. In this talk, Adam Gleave will discuss recent work by himself and collaborators scaling inverse reinforcement learning to video games, demonstrations from multiple users with differing preferences, and the very hard problem of learning from users with cognitive biases. The talk will be based on Inverse reinforcement learning for video games, Multi-task Maximum Causal Entropy Inverse Reinforcement Learning and Inferring Reward Functions from Demonstrators with Unknown Biases.

Adam is a PhD student at UC Berkeley working with Stuart Russell in the Center for Human-Compatible AI. After his talk, we will have time to discuss with him how he started working in alignment, what are the most promising approaches and ways of getting involved, and more.

There will definitely be snacks and pizza.