Inverse Reinforcement Learning

• Inverse Reinforcement Learning
• Intro
• Max Margin Formulation
• Reward Formulation

Intro

• Develop a policy based on observations of optimal behaviour
• Learning from Demonstration

Max Margin Formulation

The margin between π^* and π needs to decrease (logically)

[source: Abbeel]

• problem when big number of constraints

Reward Formulation

Get an R that satisfies:

[source: Abbeel]

• better for large domains