Metalevel Architecture of RALPH-MEA

The four execution architectures (EA) may provide conflicting decision choices, requiring some sort of arbitration scheme. This metalevel layer receives as input the decisions of each EA and uses its meta-knowledge about the quality of each EA's performance given fixed computation times. At this point it begins the meta-reasoning task of deciding whether to continue deliberation or simply output the best action choice. The latter occurs when the cost associated with the agent's limited response time and limited resources outweighs the expected utility of future computation

Continuing deliberation results in consideration of action sequences, which is the essence of planning. Although none of the knowledge types contain information about states more than two time units into the future, the Markov assumption allows concatenation of probabilities to reason about such states. In other words, the influence diagrams of the EAs can be re-used with only a change to the initial time. This can be viewed as advancing a planning window, such that to look one layer ahead means dropping the oldest layer from consideration. This augmented structure is called a dynamic influence diagram.


To return, press HOME. To go to the next document, press NEXT.