Learning Behavioural Norms for Autonomous System

This project focuses on the design and implementation of human interpretable intelligent autonomous systems. In particular the project will focus on developing and combining the following concepts and functionalities:

• Modelling of norms of behaviour and interaction
• Learning of these norms using logic-based learning, and comparison of outcomes using reinforcement or deep learning
• Incorporating reactive and pro-active behaviours
• Incorporation decision making, for example combining preferences and prospection

Concrete applications can be autonomous vehicles, health care or smart contracts.

Prospective reasoning allows hypothetical what-if scenarios while deciding what to do next. The scenarios could involve hypothetical actions (what-if this action were performed at this time?) or hypothetical observations (what if this were observed at this time?). By `simulating’ the possible evolutions of the current state, the system should be able to determine the consequences of an action plan before executing it, and hence to choose a course of actions that will lead to the best long-term outcome.

The project will start with a concrete existing framework, LPS – Logic Production Systems, a logic-based paradigm incorporating reactivity, proposed by Kowalski and Sadri. The framework is based on abduction and combines reactive rules, logic programs and a database together with a causal theory that specifies how database states make transitions. An implementation of LPS exists in ASP (Answer Set Programming) that can provide the basis of this project.

The explicit formalisation of the norms of behaviour and reactivity will facilitate transparency and ease of understanding and can form the basis of trust in the learned norms. The possibility of simulating behaviour in ASP allows further confidence. LPS also has an implementation in JavaScript that lends itself well to visualisation.

For the learning part the research can explore inductive learning techniques, for example using ILASP (Inductive Learning of Answer Set Programs) which is a recently developed system for machine learning of ASP programs from examples.

The project can further explore the identification and formlisation of formal properties that would encourage trust in the system. The use of logic throughout for modelling and the use of model generation paradigm of ASP can be exploited to prove and analyse such properties.

Project ID

STAI-CDT-2020-IC-13

Supervisor