Explanation-aware Computing 

Abstract: A Markov Decision Process (MDP) policy presents, for each state, an action, which preferably maximizes the expected reward accrual over time. In this paper, we present a novel system that generates, in real time, natural language explanations of the optimal action, recommended by an MDP while the user interacts with the MDP policy. We rely on natural language explanations in order to build trust between the user and the explanation system, leveraging existing research in psychology in order to generate salient explanations for the end user. Our explanation system is designed for portability between domains and uses a combination of domain specific and domain independent techniques. The system automatically extracts implicit knowledge from an MDP model and accompanying policy. This policy-based explanation system can be ported between applications without additional effort by knowledge engineers or model builders. Our system separates domain-specific data from the explanation logic, allowing for a robust system capable of incremental upgrades. Domain-specific explanations are generated through case-based explanation techniques specific to the domain and a knowledge base of concept mappings for our natural language model.

Abstract: In order to assist a power plant operator to face unusual situations, we have developed an intelligent assistant that explains the suggested commands generated by an MDP-based planning system. This assistant provides the trainee a better understanding of the recommended actions to later generalize them to similar situations. In a first stage, built-in explanations are predefined by a domain expert and encapsulated within explanation units. When the operator takes an incorrect action, an explanation is automatically generated. A controlled user study in this stage showed that explanations have a positive impact on learning. In a second stage, we are developing an automatic explanation generation mechanism based on a factored representation of the decision model used by the planning system. As part of this stage, we describe an algorithm to select a relevant variable, which is a key component of the explanations defined by the expert.

Abstract: In this paper a representative example is chosen that is meant be fairly simple for illustrating the point that in a very common kind of instance, argument and explanation are mixed in together in a text of discourse. The example is a short text found on the Internet that explains to the reader how to attach a flagpole bracket to the vinyl siding on the side of your house. The example uses practical reasoning (goal-directed reasoning) of a kind widely studied in AI and logic. While the text appears to be mainly a “how-to” explanation, it also contains argumentation woven into it, as shown by applying argumentation schemes (defeasible argument structures) representing common forms of argument. The problem is one of distinguishing between explanation and

Abstract: This work focuses on the socio-technical aspects of artificial intelligence, namely how (specific types of) intelligent systems function in human workplace environments. The goal is first to get a better understanding of human needs and expectations when it comes to interaction with intelligent systems, and then to make use of the understanding gained in the process of designing and implementing such systems.The work presented focusses on a specific problem in developing intelligent systems, namely how the artefacts to be developed can fit smoothly into existing socio-cultural settings. To achieve this, we make use of theories from the fields of organisational psychology, sociology, and linguistics. This is in line with approaches commonly found in AI. However, most of the existing work deals with individual aspects, like how to mimic the behaviour or emulate methods of reasoning found in humans, whereas our work centers around the social aspect. Therefore, we base our work on theories that have not yet gained much attention in intelligent systems design. To be able to make them fruitful for intelligent systems research and development, we have to adapt them to the specific settings, and we have to transform them to suit the practical problems at hand.The specific theoretical frameworks we draw on are first and foremost activity theory and to a lesser degree semiotics. Activity theory builds on the works of Leont'ev. It is a descriptive tool to help understand the unity of consciousness and activity. Its focus lies on individual and collective work practise. One of its strengths, and the primary reason for its value in AI development, is the ability to identify the role of material artefacts in the work process. Halliday's systemic functional theory of language (SFL) is a social semiotic theory that sets out from the assumption that humans are social beings that are inclined to interact and that this interaction is inherently multimodal. We interact not just with each other, but with our own constructions and with our natural world. These are all different forms of interaction, but they are all sign processes.Due to the obvious time and spatial constraints, we cannot address all of the challenges that we face when building intelligent artefacts. In reducing the scope of the thesis, we have focused on the problem of explanation, and here in particular the problem of explanation from a user perspective. By putting social theories to work in the field of artificial intelligence, we show that results from other fields can be beneficial in understanding what explanatory capabilities are needed for a given intelligent system, and to ascertain in which situations an explanation should be delivered. Besides lessons learned in knowledge based system development, the most important input comes from activity theory.The second focus is the challenge of contextualisation. Here we show that work in other scientific fields can be put to use in the development of context aware or ambient intelligent systems. Again, we draw on results from activity theory and combine this with insights from semiotics.Explanations are themselves contextual, so the third challenge is to explore the space spanned by the two dimensions ability to explain and contextualisation. Again, activity theory is beneficial in resolving this issue.The different theoretical considerations have also led to some practical approaches. Working with activity theory helps to better understand what the relevant contextual aspects of a given application are and helps to develop models of context which are both grounded in the tradition of context aware systems design and are plausible from a cognitive point of view.Insights from an analysis of research in the knowledge based system area and activity theory have further lead to the amendment of a toolbox for requirements engineering, so called problem frames. New problem frames that target explanation aware ambient intelligent systems are presented. This is supplemented with work looking at the design of an actual system after the requirements have been elicited and specified. Thus, the socio-technical perspective on explanations is coupled with work that addresses knowledge representation issues, namely how to model sufficient knowledge to be able to deliver explanations.