RALPH-MEA: A Real-Time, Decision-Theoretic Agent Architecture
Abstract: This dissertation describes the RALPH-MEA agent architecture which uses decision theory combined with real-time control for decision-making in complex domains. In order to achieve the conflicting goals of an accurate representation and a fast decision cycle, several novel techniques are introduced.
Multiple execution architectures are four implementations of the agent function, a function that receives percepts from the environment as input and outputs an action choice. The four execution architectures (EAs) are defined by the different knowledge types that each uses. Depending on the domain and agent capabilities, each EA has different advantages. For example, a reactive, "if (condition) then (action)", production rule system will generally allow fast reactions, while a deliberative, decision-theoretic system will be slow but accurate and easily programmed with new knowledge and goals. A metalevel algorithm to combine the results of multiple EAs is given, and a decision-theoretic representation of the EAs as "extended influence diagrams" is defined.
Knowledge compilation is used to convert knowledge of one type to another. For example, the knowledge used by a decision-theoretic system (e.g., probabilities and utilities of outcome states) can be converted into knowledge used by a condition-action rule systems. A viable strategy is to acquire knowledge in one form and then to use knowledge compilation to convert the knowledge into the most efficiently executable form.
A view of decision-theoretic planning is also presented. Utilizing decision theory for planning facilitates the handling of uncertainty and multiple objectives. However, because of the high complexity of such planning, control of planning becomes a critical issue. Metalevel control of planning computes the value of information of planning to compare to the utility of executing the current default plan.