Posts

• Oct 20, 2020 reinforcement-learning  inverse-rl  gan 

  Infer Rewards by Observing an Expert

  Designing a reward function for a Reinforcement Learning(RL) task can prove notoriously difficult. No formula exists to guide reward function design – a desirable reward function is often realised through trial and error. However, that won’t always work.

  Inverse Reinforcement Learning (iRL) seeks to provide an alternative to hand-engineered reward functions by recovering a suitable reward function from demonstrations of desired behaviour. This post details how iRL works and guides on the implementation of an adversarial iRL algorithm.