A Taxonomy of RL Algorithms

taxonomy: 분류(학)

Model-Free vs Model-Based RL

Model-Based RL

Agent has access to a model of the environment

advantages

• allows the agent to plan e.g. AlphaZero

disadvantages

• models are usually not available (especially in real world)
• if we want to use model-based we need to learn the model from experience and create a predictive representation of the environment. This creates several challenges
  1. bias can be exploited
  2. needs lots of time and compute but does not guarentee success
  3. etc …

Model-Free RL

Agent does not access to a model of the environment

In 2018 it was more popular and have been more extensively developed and tested than model-based methods

advantages