Recently I came across an interesting comment about reinforcement learning and couldn't really understand it. It's something like:

Reinforcement learning is the only type of machine learning where you are allowed to blatantly train on test set.

Is it correct/accurate? Why?


The statement has some sense but it is not entirely right. It depends on the task that you want to achieve. Most reinforcement learning scenarios are applied to games, which means the task you want to achieve is usually to win the game, where there is only a single environment, you can just train the agent on that single environment (Eg. Breakout). The end result is that the agent is able to proficiently "win" in this environment, as a by product of reward maximisation. Some references in which reinforcement learning is applied in this case is in the atari games and alphaGo

There are certain scenarios in which you might want to have your agent generalise to "test" environments. In this case, the agent is first trained on some training levels, before they are placed in a test environment to see how well they perform. This is quite common in many game scenarios as well, where the environment need not be the same as you "progress". There is ongoing research in studying how reinforcement learning agents generalise to unseen environments.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.