Whence reward

How to define reward in a reinforcement learning framework?

• Programming

1. Coding: translate the goals of behaviour into reward values, takes states outputs rewards
2. Human-in-the-loop: source of reward is person, non-stationary reward

• Example

1. Mimic reward: copy the given reward
2. Inverse reinforcement learning: learner would figure out what rewards the trainer must have been maximizing that makes this behaviour optimal

• Indirect approaches, optimization

1. Evolutionary optimizationH high-level behaviour we can create a score for, and optimization would search for reward to encourage the behaviour
2. Meta RL: learning at evolutionary level that creates better ways of learning at the individual level