| Abstract
|
to respond effectively in real time, an ag … to respond effectively in real time, an agent needs to be able to dealeffectively with differing availability of information. in particular,although the agent ideally would like to be able to diagnose one or moreparticular faults that may be present at a given time, a deadline mayprevent it from doing so because prior to the deadline it is required toact or face a catastrophic outcome. thus, an agent should have availableto it a set of actions some of which are appropriate, but not optimal, forlarge classes of faults and some of which are appropriate for specificfaults.tests are available to help the agent diagnose a fault; given this frameworkdecision trees would provide one possible approach to this diagnosis problem.however this work shows that the information-theoretic heuristic commonly used to structure decision trees is not ideal in a real time domain becauseit concerns itself with reaching a leaf node as fast as possible, not withthe value of the actions it might obtain along the way. my thesis presentsan alternative heuristic, the action-based heuristic, which may be used tostructure a decision tree; the resulting structure together with theactions themselves is called an action-based hierarchy.the approach is validated in several ways. first a complexity analysis isundertaken to show that the complexity of the structuring algorithm is notprohibitive. then some theoretical results are given; this is followed byexperiments both with abstract inputs and inputs from a real-time domain,surgical intensive care unit patient monitoring. the thesis concludes witha description of an implementation of these ideas in a system known as react. of these ideas in a system known as react.
|
