Explaining the Space of Plans

Abstract

In today s technological development, more and more technical systems will take action decisions traditionally taken by humans. Where such decisions directly affect economic value, or even human lives, it is crucial for human users to be able to check the decision rationales. The ability to explain decisions suggested by a computer system is, therefore, increasingly recognized as crucial to the success and practicality of such systems. The objective of the proposed project is to establish explanation facilities for AI Planning, in terms of a Q&A process. Like other model-based control methodologies -- in difference to model-free control through ML methods -- AI Planning lends itself naturally to explanation, as the reasoning process behind the suggested decisions is explicit and can, in principle, be checked by the human user. The difficulty then lies in actually making such reasoning -- enumerating vast spaces of alternate decisions -- amenable to human users. Our central thesis is that this can be naturally done in terms of explaining the space of plans, pointing out the most relevant plan properties and their dependencies. For example, in a robotics-planning application supporting underwater vehicle control, given a plan of action suggested by the system, the user may ask ``why not cover objective x? ; the answer could be ``covering x would either require excessive time or energy, or result in abandoning y or z . Such an answer would be derived by considering the space of relevant plan properties -- objectives covered, energy consumption, time – and examining their relative behavior within the space of plans. A user question then translates into an enforced plan property p (cover objective x), and the answer is given in terms of the most relevant (detrimental/non-trivial) plan properties q entailed by p. Given such an answer, the user may choose to ask a different question, or -- if p is desirable despite q -- request an alternative plan satisfying p.

Document Details

Document Type
DoD Grant Award
Publication Date
Jun 11, 2018
Source ID
FA95501810245

Entities

People

  • Daniele Magazzeni

Organizations

  • Air Force Office of Scientific Research
  • United States Air Force

Tags

Readers

  • Software Engineering
  • Team-Based Human-Centered Cognitive Task Decision Making and Information Performance.
  • Theoretical Analysis.

Technology Areas

  • AI & ML
  • AI & ML - DoD AI Strategy
  • AI & ML - Machine Learning Algorithms
  • Autonomy
  • Autonomy - Human-Robot Interaction
  • Space