Terminology
Environment is fully observable if agent's sensors can detect all information required to make optimal decision at any point in time i.e. Sensors can always perceive entire state of environment.Â
If agent requires memory, then environment is only partially observable. The sensors can only perceive fraction of the environment's state at any given moment.
Memory of past measurements provides additional information in a partially observable environment e.g cards that have already been discarded in a game of Blackjack.
Environment is deterministic if agent's actions uniquely determine the outcome e.g. chess - pieces move according to rules, always the same outcome to the same move.Â
If environment has random element, is stochastic e.g. dice games.Â
Discrete - finite number of possible states / finite number of possible actions. Continuous - infinite number of states / possible actions.Â
Benign -Â environment has no objective of its own to oppose the agent's. Adversarial -Â environment has own objective e.g. opposed games. Harder to take good actions when opposed.Â
