Human-aware PGMs and Probabilistic Inference via Lifted Model Reconciliation (HAPPI)

KI Starter Project

• Introduction

  In AI-based systems, models are often learned from data. Although these models can also contain knowledge from experts, they are often no longer easy to explain using the information learned from the data. The HAPPI project addresses this by researching methods for reconciling such learned models with the expectations of experts and enabling humans and systems to explain their knowledge gaps to each other such that, in the long term, these systems can explain their suggestions to users when used to provide support, for example, in the medical field or within (digital) humanities.

  Specifically, probabilistic models and inference are fundamental building blocks of knowledge-driven artificial intelligence but unfortunately hard to interpret, even if using techniques such as lifting to compactify a model. In human-aware planning, model reconciliation is used to explain differences between a proposed and an ex- pected plan. The goal of this project is to investigate model reconciliation in combination with lifting for proba- bilistic inference, with applications in the area of medicine and digital humanities. Specifically, the contributions are two-fold: (a) hierarchical lifting to build a hierarchy of models to use as a basis for model reconciliation and (b) model reconciliation in the form of model update computations between propositional and lifted models as well as adaptation to user feedback. The project builds upon my existing expertise in lifted inference and expands on it by researching the combination of human-awareness, lifting, and probabilistic inference.

• Project members

  • Tanya Braun, project lead
  • N.N., research associate
  • N.N., student assistant

• Research Goals

  The goal of the project is to research model reconciliation for explanations in probabilistic inference with the help of lifting. To ground both the user’s and the system’s model in a shared world, we assume an expert knowledge model for the user that is augmented with available data through standard machine-learning techniques for PGMs. The system’s model, thus, contains more information extracted from the data that is otherwise unavailable to the user. Lifting then has the purpose to compactify the system’s model for enabling real-time answers as well as for providing a basis for reconciliations through generalisation as well as specialisation. To this end, the project consists of two major parts: (i) We will research a hierarchical lifting for supporting different levels of generalisation as a basis for model reconciliation. (ii) We will research model reconciliation for PGMs based on hierarchical lifting.

• Publications

  to appear