Understanding the reasoning behind automated planning is essential to building confidence in autonomous systems. Multi-objective probabilistic planning has been used in many domains for synthesizing behavior for autonomous systems in uncertain environment, where the systems must optimize their behavior for multiple qualities of concern (referred to as quality attributes). It is important for humans overseeing or interacting with those systems to understand (i) what the planning solutions are, and what quality-attribute properties those solutions have, and (ii) why the planning processes selected those solutions, in terms of the quality attributes of concern.
Multi-objective probabilistic planning can be formulated as multi-objective Markov decision processes (MDPs). However, understanding the planning as discussed above is challenging. The sequential and nondeterministic nature of policies makes it difficult for humans to make sense of MDP-based planning solutions. Moreover, due to opaque evaluation of quality attributes and complex preference and value-tradeoff analyses of conflicting quality objectives, it is challenging for humans to understand why a particular solution policy is generated.
I propose to promote transparency and understandability of multi-objective MDP planning, by automatically explaining and justifying solution policies on the basis of the corresponding quality attributes and their optimization objectives. My approach comprises four elements: (i) Explainable planning language, which preserves the semantics of quality attributes, their evaluation models, and their utility models in the planning problem formulation. (ii) Policy summarization for describing a solution policy concisely. (iii) Contrastive quality-attribute-based justification for arguing why a solution policy is best among available alternatives, with respect to the quality attributes. Lastly, (iv) Quality-attribute evaluation explanation for describing the determining factors of the quality-attribute properties of a policy.
David Garlan (Co-Chair)
Reid Simmons (Co-chair, Robotics)
Bashar Nuseibeh (The Open University)