Third edition of Artificial Intelligence: foundations of computational agents, Cambridge University Press, 2023 is now available (including the full text). Effect Uncertainty

In some cases an agent knows the effect of an action. That is, given a state and an action, it can accurately predict the state resulting from carrying out that action in that state. For example, an agent interacting with a file system may be able to predict the effect of deleting a file given the state of the file system. In many cases, it is difficult to predict the effect of an action, and the best an agent can do is to have a probability distribution over the effects. For example, a person may not know the effect of calling his dog, even if he knew the state of the dog, but, based on experience, he has some idea of what it will do. The dog owner may even have some idea of what another dog, that he has never seen before, will do if he calls it.

The effect uncertainty dimension is that the dynamics can be

  • deterministic--when the state resulting from an action is determined by an action and the prior state or
  • stochastic--when there is only a probability distribution over the resulting states.

This dimension only makes sense when the world is fully observable. If the world is partially observable, a stochastic system can be modeled as a deterministic system where the effect of an action depends on some unobserved feature. It is a separate dimension because many of the frameworks developed are for the fully observable, stochastic action case.

Planning with deterministic actions is considered in Chapter 8. Planning with stochastic actions and with partially observable domains is considered in Chapter 9.