I have recently solved the Cartpole problem using double deep Q-learning. When I saw how the agent was doing, it used to go right every time, never left, and it did similar actions all the time.

Did the model overfit the environment? It seems that the agent just memorized the environment.

What are the common techniques to prevent the agent to overfit like that? Is that a common problem?

  • 1
    $\begingroup$ In supervised learning, overfitting generally means that you learn the training data, but don't perform well on the test data. The concept of overfitting in reinforcement learning makes sense in the case if your training environment is different than your evaluation/test environment. However, if you train your RL agent in the same exact environment where you want it to behave, you actually want it to "overfit", i.e. learn the environment dynamics. See this answer for more details. $\endgroup$
    – nbro
    Commented Jun 2, 2020 at 15:18


You must log in to answer this question.