Hands-On Intelligent Agents with OpenAI Gym

Reward clipping

Different problems or environments provide different ranges of reward values. For example, we saw in the previous chapter that in the Mountain Car v0 environment, the agent receives a reward of -1 for every time step until episode termination, no matter which way the agent moves the car. In the Cart Pole v0 environment, the agent receives a reward of +1 for every time step until episode termination. In Atari game environments like MS Pac-Man, if the agent eats a single ghost, it will receive a reward of up to +1,600. We can start to see how the magnitudes of the reward as well as the occasion of the reward varies widely across different environments and learning problems. If our deep Q-learner agent algorithm has to solve ...