Skip to content

Using RL with Gymnasium

Programming / ai / reinforcement learning

Custom Env

Gymnasium main concepts

  • Observation Space: Set of possible state that agent can observe in the environment
  • Action Space: Set of actions that agent cant take in environment
  • Episode: A complete run through the environment from initial state until terminate state is reach, each episode is composed if a sequence of states, actions and rewards
  • Wrapper: A tool in GYM that allow modify an environment behavior without changing its code, for example and time constrain and action masking
  • Benchmark: Help to compare between different RL algorithm

Observation space

What information the environment gives the agent

Action space

What actions the agent is allowed to take

  • next_state: the observation after taking the action
  • reward: the reward after tacking the action
  • terminated: boolean, true if episode ended
  • truncated: boolean , true if the episode end by early truncation (time limit reached)
  • info: a dictionary contain additional environment information (for example in atari game it's hold user lives)

Reference