RI: Small: Expressive Reasoning and Learning about Actions under Uncertainty via Probabilistic Extension of Action Language

Arizona State University, Scottsdale AZ

Investigators

Abstract

Automated reasoning about dynamic worlds is an important capability for robust intelligent systems. Action languages allow for the description of actions and their effects in dynamic domains in a way that is based on natural language but sufficiently formal for modeling in knowledge-based systems. Today's action languages do not easily allow such systems to account for the probability and uncertainty necessary to model human-like commonsense reasoning. Existing action languages also assume full specification of a system in advance of one-shot execution of the logic program, which does not easily operate with continuous streams of data. This project will develop an action language based on the mathematical foundation that combines logic and probability. The research will join the representation and reasoning advantages of logical AI to the advantages in statistical AI to compute and learn quantitative specifications from data. The new action language will jointly address commonsense reasoning and learning about actions in uncertain dynamic domains. Such a system allows us to scrutinize and understand the system behavior, which is vital to the design of systems that are explainable and interpretable. The project is to design and implement a novel action language that is highly expressive for modeling various aspects of dynamic systems under uncertainty and which applies to knowledge-rich diagnosis and stream reasoning. The formalism will be built upon a recent probabilistic extension of answer set programs, called LPMLN, which incorporates the weight scheme of Markov Logic into the language of answer set programming. The formalism will enable probabilistic diagnostic reasoning and counterfactual reasoning about dynamic domains. Inference and learning methods for the probabilistic action language will be derived from the methods in logic programming and statistical relational learning. The framework will be further extended to integrate reasoning over observations given as streams of data. The methods produced will be useful for several applications that require integration of knowledge representation and other areas, such as robotics and autonomous systems. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.