AI seminar: Extending environment design to a multiagent setting

Speaker: Lachlan Dufton, University of Waterloo

Policy teaching through environment design is a new framework that allows an interested party to modify an agent's environment by providing incentives, in order to achieve a desired behaviour. By observing changes in the agent's behaviour, the interested party can learn the agent's preferences with active indirect elicitation. Presently, this framework deals with a single agent setting.

In this talk I will present an overview of environment design, policy teaching and active indirect elicitation. I will then present my current results for extending this framework to a multiagent setting.