This is a wide area of Artifical Intelligence that builds upon the large body of research and work that fall under what is currently referred to as GOFAI: Good Old Fashion AI. It includes a range of teachiniques for ensuring safety, trustworthiness, robustness and interpretability of AI systems. They range from logic, model-based automated planning, explanation generation, reasoning under uncertanty and incomplete knowledge, diagnosis, and learning. We tackle open research challenges in each of these aspects of symbolic AI with the objective of developing algorithms and systems that are general, scalable and effective in their task, but also based on a rigourous semantics.

Abductive inference

Abductive inference is one of the three forms of human reasoning, the other two be deductive and inductive. A larg area of research in computationa logic has seen the development of a range of framework, algorithms and systems for performing abductive inference. We haev been focusing our efforts on two aspects of abductive inference: distributed abdution over mulit-agent systems and extension of abductive inference with probabilistic inference. Specifically, we have developed theory, implementation and application of a general distributed abductive reasoning framework called DARE, and a novel probabilistic abductive framework based on the distribution semantics for normal logic programs that handles negation as failure and integrity constraints in the form of denials. For application of abductive inference see our Application research theme.

For a list of our published work on this topic, please visit our publications website.

Planning

Planning is a well-established area of AI. Open research challenges include (i) developing heuristics that make the search more efficient and (ii) combining search startegies with additional exploration mechanisms to enable the search to escape local minima and plateaus of the heuristic functions. Modern classical planners usually rely on heuristic forward search with methos for learning domain-specific heuristics limiting their transferability from one task to another. Domain-independent heuristic functions tend to be encode by human experts. We are interested in exploring ways in which machine learning techniques can be used to improve the efficiecy of model-based planning search strategies. Specifically, develop machine learning solutions to automate the design of multi-technique search algorithm by learning planning strategies through policy gradient methods, and to learn effective domain-independent planning heuristics to help generalize the applicability of planning to a variety of problems including for instance that of dynamic (collective) adaptive systems.

For a list of our published work on this topic, please visit our publications website.

Probabilistic Structured Inference

Combining probabilstic inference and paramater learning with logic-based inference is one of the areas of AI targetted at solving tasks where reasoning in the presence of incomplete knowledge, uncertanty and noise are key dominant factors. As logic based inference can be of three types (demduction, abduction and induction) we explore the integration of probabilistic inference with each of three different forms of human inference. We have made significate contribution on the development of (1) probabilistic deductive inference, in the context of asnwer set programming by proposing the first solver for Probabilsitic answer set programming, called PASOCS, (2) probabilistic abduction and (3) probabilistic induction. These areas are research present still many open challenges with a range of open PhD research topics.

For a list of our published work on this topic, please visit our publications website.

Symbolic Machine Learning

Learning interpretable knowledge from data is one of the main challenges of AI. There has been a growing interest in Symbolic Machine Learning, a field of Machine Learning that aims at developing algorithms and systems for learning logic-based programs that explain labelled data within the context of some given background knowledge. Our goal to develop novel, effective, and scalable symbolic machine learning algorithms and systems that can provide proof guarantees, robustness to noise in the data and customisable through domain-driven optimisation criteria. Our family of symbolic machine leanring systems, including in particular the state-of-the-art systems ILASP and FastLAS, boasts a number of advanced features. They can support the learing of non-monotonic and non-deterministic programs, programs that capture preference reasoning through weak constraints, and programs that include domain-specific hard constraints. All these features make them suitable to tackle real-world problems including learning common-sense interpretable knowledge from unstructured data.

For a list of our published work on this topic, please visit our publications website.