Algos Expeditus

There are 4 key aspects of Reinforcement Learning :

1. Optimization
2. Delayed Consequences
3. Exploration
4. Generalization

1. Optimization

There are different types of decision to be made, and we want to make the best possible decisions, or atleast good decisions

2. Delayed Consequences

Let’s say you’re playing Mario. In this case, the agent is Mario, and if we let him take mushroom in an earlier level, it will help him later on in the game. This also gives rise to something known as the credit assignment problem, i.e., how do we know for sure that the relationship between the decisions we took in the past, and the outcomes of the future have a causal relationship.

3. Exploration

Learning by making decisions. We also have to consider about censored data, which means the agent only gets rewards for the decision it makes

4. Generalization

Policy is mapping from past experience to action. You’re given an idea of how the world works