Ralph: Planning

Description

• Planning is the process of projecting forward from the current state to predict possible outcomes
• In simple domains, planning is unnecssary, because the maximum expected utility can be calculated directly
• In more complex domains, in order to determine the utility of the current action, we must consider possible subsequent sequences of actions
  • The Decision-Theoretic and Goal-Based execution architectures are capable of doing this explicitly
  • The Action-Utility and Condition-Action architectures must do this indirectly
• Decision-theoretic planning is concerned with determining what sequence of action has the highest expected utility, then executing the first action of that sequence
• Evaluating sequences of actions is exponentially more expensive than evaluating one action, so successful planning must involve additional control; RALPH uses replanning control: if, while planning, the utility of replanning is greater than the utility of planning, replanning will be selected instead of continued planning