Jump to the RAPLH-MEA Overview page.
Jump to the Architectural Details Menu page.
Jump to the Home page.

Architectural Details

The RAPLH-MEA architecture is an attempt to use decision theory in an actual implementation in a real-time domain. The authors explain how the architecture uses different types of knowledge as described in this figure:

An agent sits in a current state from which it can take several actions which lead it to some next state. Each one of these next states will have a utility. The maximum expected utility MEU principle can then be used to determine which is the best action.

The different types of knowledge, denoted by arrows in the figure, are:

A: Information about current state.
B: Effects of actions.
C: Utilities of projected state.
D: Best action choice directly from current state.
E: Utilities of (current state,action) pairs.
F: Best action choice directly from projected state.

The architecture then multiple the multiple execution architectures MEA as those modules that use each one of the possible subsets of types of knowledge. These are:

DT: Decision Theoretic EA, uses A B C MEU.
GB: Goal-Based EA, uses A B F.
AU: Action-Utility EA, uses A E MEU.
CA: Condition-Action EA, uses A D.

The architecture runs each one of these EA in parallel and their results, when they have them (all this happens asynchronously) go to a metalevel. The metalevel has knowledge about the performance of each EA given a certain amount of computation time. Using this information it determines the value of further computation and stops computing (i.e. tell the MEA to give it up) when the expected value falls below zero.

The architecture also provides support for planning and replanning. It does this by computing a plan and, at each step, executing the first step of the plan. Since the results of the agent's actions might coincide with what was expected the agent might eventually find itself off the plan. However, the architecture only decides to replan when the utility of replanning is greater than the utility of executing the plan. There is a metalevel control that determines when and how much replanning should be done. THis is done by viewing replanning as a computational action that can be integrated into the same decision cycle as base-level actions (i.e actions that directly affect the external world). On each cycle an action is selected, this action might be a base-level action or a replanning action.