: In this paper, we consider the problem of improving the goalachievement performance of an agent acting in a partially observable, dynamic environment, which may or may not know all events that can happen in that environment. Such an agent cannot reliably predict future events and observations. However, given event models for some of the events that occur, it can improve its predictions of future states by conducting an explanation process that reveals unobserved events and facts that were true at some time in the past. In this paper we describe the DISCOVERHISTORY algorithm for discovering an explanation for a series of observations in the form of an event history and a set of assumptions about the initial state. When knowledge of one or more event models is not present, we claim that the capability to learn these unknown event models would improve performance of an agent using DISCOVERHISTORY, and provide experimental evidence to support this claim. We provide a description of this problem and suggest how the DISCOVERHISTORY algorithm can be used in that learning process.
: We consider the problem of automated planning in partiallyobservable dynamic environments, where exogenous events that cannot be directly observed affect the state of the world. In these environments, a planner?s knowledge of the world is limited, and state transitions can be both ambiguous and difficult to predict due to that lack of knowledge. We describe a new formalism and new algorithms that enable a planner to proactively expand its knowledge of the environment during planning and execution, by modeling the exogenous events that can occur and forming explanations that reveal information about the world. We have implemented our new algorithms in a variant of the well-known SHOP2 planner that can replan when a failure occurs during plan execution. We have conducted an ablation study in two planning domains to examine the effects of explanation on execution. The results demonstrate that our algorithm successfully increases the performance of an agent using it in two planning domains. This improvement results from the agent having increased knowledge of the environment which allows it to more accurately predict future events and ultimately make better plans.
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.
Social psychology research on teams has shown that team performance improves when team members have a so-called shared mental model (SMM). We observe that a decision support system (DSS) and its user can be viewed as a team whose task is to make a (good) decision. Therefore we argue that it is important for good user-DSS team performance that they have a SMM. In this paper we make precise what it means for a user and DSS to have a SMM. This can be used as the basis for analyzing to what extent a DSS supports the achievement of a SMM, or to design a DSS such that it does this. We apply our framework to a particular negotiation support system, using it to analyze to what extent this DSS facilitates the achievement of a SMM. We propose that the addition of explanation facilities to the DSS can improve sharedness in the user-DSS team.
According to various embodiments of the invention, a relay is suitably formed to exhibit an open state and a closed state. The relay is operated by providing a cantilever sensitive to magnetic fields such that the cantilever exhibits a first state corresponding to the open state of the relay and a second state corresponding to the closed state of the relay. A first magnetic field may be provided to induce a magnetic torque in the cantilever, and the cantilever may be switched between the first state and the second state with a second magnetic field that may be generated by, for example, a conductor formed on a substrate with the relay.
. We implemented a generic speech-based dialogue shell that can be configured for and applied to domain-specific dialogue applications. A toolbox for ontology-based dialogue engineering provides a technical solution for the two challenges of engineering domain extensions for new question and answer possibilities and debugging functional modules. In this paper, we address the process of debugging and maintaining rule-based input interpretation modules. While supporting a rapid implementation cycle until the dialogue systems works robustly for a new domain (e.g., the dialogue-based retrieval of medical images), production rules for input interpretation have to be monitored, configured, and maintained. We implemented a special graphical user interface to monitor and explain reasoning processes for the input interpretation phase of multimodal dialogue systems. A particular challenge was the presentation of the software system’s ontology-based interaction rules in a way that they were accessible to and editable for humans for maintenance, and, at the same time, allowed a real-time monitoring of their application in the running dialogue system.
We implemented a generic dialogue shell that can be configured for and applied to domain-specific dialogue applications. The dialogue system works robustly for a new domain when the application backend can automatically infer previously unknown knowledge (facts) and provide explanations for the inference steps involved. For this purpose, we employ URDF, a query engine for uncertain and potentially inconsistent RDF knowledge bases. URDF supports rule-based, first-order predicate logic as used in OWL-Lite and OWL-DL, with simple and effective top-down reasoning capabilities. This mechanism also generates explanation graphs. These graphs can then be displayed in the GUI of the dialogue shell and help the user understand the underlying reasoning processes. We believe that proper explanations are a main factor for increasing the level of user trust in end-to-end human-computer interaction systems.
This paper presents an approach for providing explanation to the intelligent diagnosis and monitoring of business workflows based on operation data in the form of temporal log date. The representation of workflow related case knowledge in this research using graphs is explained. Workflow cases are represented in terms of events and their corresponding temporal relationships. The matching and CBR retrieval mechanisms used in this research are explained and the architecture of an integrated intelligent monitoring system is shown. The paper contains an illustration and evaluation of the approach based on experiments on real data from a university quality assurance exam moderation system. It is shown that a graph matching based similarity measure is capable to diagnose problems within business workflows and the results of this reasoning process can be explained to workflow managers and users with the use of graph visualisation techniques.
The provision of explanations in knowledge based systems has a long tradition in computer science. Regarding expert systems, several interesting contributions were proposed, but an overall method for explanation generation is missing. Furthermore, various approaches have never been realised or have never been followed up with respect to current technologies. Explanation-Aware Software Design (EASD) aims at making software systems smarter in interactions with their users. The long-term goal is to develop methods and tools for engineering and improving such capabilities. In this paper we present the idea of EASD including an abstract model for explanation generation. We describe in detail the realisation of that approach for the semantic search engine KOIOS using Semantic Web technology.
Data mining methods build patterns or models. When presenting these, all or part of the result needs to be explained to the user in order to be understandable and for increasing the user acceptance of the patterns. In doing that, a variety of dimensions in the Mining and Analysis Continuum of Explaining (MACE) needs to be considered, e.g., from concrete to more abstract explanations. This paper discusses the application of the MACE in the context of social software. We consider applications of the proposed approaches in three social software systems, and show how the data mining results can seamlessly be analysed on the presented continuous dimensions and levels.
The main advantage of machine learning algorithms that learn simple symbolic models is in their capability to trivially provide justifications for their decisions. However, there is no guarantee that these justifications will be understood by experts and other users. Induced models are often strange to the domain experts as they understand the problem in a different way. We suggest the use of argument based machine learning (ABML) to deal with this problem. This approach combines machine learning with explanations provided by domain experts. An ABML method is required to learn a model that correctly predicts learning examples and is consistent with the provided explanations. The present paper describes an application of ABML to learning a complex chess concept of an attack on the castled king. The explanation power of the learned model for this concept is especially important, as it will be used in a chess tutoring application.
This paper provides a study of the theoretical properties of Most Relevant Explanation (MRE) [12]. The study shows that MRE defines an implicit soft relevance measure that enables automatic pruning of less relevant or irrelevant variables when generating explanations. The measure also allows MRE to capture the intuitive phenomenon of explaining away encoded in Bayesian networks. Furthermore, we show that the solution space of MRE has a special lattice structure which yields interesting dominance relations among the candidate solutions.
We have developed an automatic explanation generation mechanism for an intelligent assistant. Based on a Markov decision processes (MDP), the assistant determines the optimal action for each state, guiding the user in the operation of an industrial plant. When the user makes an error, the assistant generates an explanation composed of 3 main parts: (i) the recommended action in the current situation; (ii) a graphical representation of the process highlighting the relevant variable; (iii) a verbal explanation. To generate the explanations we combine several knowledge sources, including the MDP and a frame system with information of the components, actions and variables in the process. In this paper we present an evaluation of the explanations generated by our system. A panel of experts compared the explanations given by the system to those given by a human, in terms of: coherence, contents, organization, and precision. We consider 15 different explanations for a variety of situations in the power plan domain, and asked 6 different external domain experts to evaluate them. The results are between “good” and “excellent” in average for all the aspects and for all the experts. We consider that these are encouraging results, as we are comparing the explanations generated automatically against those given by an experienced domain expert, a very high standard.
Applications deployed on cyber-infrastructures often rely on multiple data sources and distributed compute resources to access, process, and derive results. When application results are maps, it is possible that non-intentional imperfections can get introduced into the map generation processes because of several reasons including the use of low quality datasets, use of data filtering techniques incompatible for the kind of map to be generated, or even the use of inappropriate mapping parameters, e.g., low-resolution gridding parameters. Without some means for accessing and visualizing the provenance associated with map generation processes, i.e., metadata about information sources and methods used to derive the map, it may be impossible for most scientists to discern whether or not a map is of a required quality. Probe-It! is a tool that provides provenance visualization for results from cyber-infrastructure-based applications including maps. In this paper, we describe a quantitative user study on how Probe-It! can help scientists discriminate between high and low quality contour maps. The study had the participation of twenty active scientists from five domains with different levels of expertise with regards to gravity data and GIS. The study demonstrates that only a very small percentage of the scientists can identify imperfections using maps without the help of knowledge provenance. The study also demonstrates that most scientists, whether GIS experts, subject matter experts (i.e., experts on gravity data maps) or not, can identify and explain several kinds of map imperfections when using provenance to inspect maps.
Across many fields involving complex computing, software systems are being augmented with workflow logging functionality. The log data can be effectively organized using declarative structured languages such as OWL; however, such declarative encodings alone are not enough to facilitate understandable workflow systems with high quality explanation. In this paper, we present our approach for visually explaining OWL-encoded workflow logs for complex systems, which includes the following steps: (i) identifying and normalizing provenance in workflow logs using the provenance interlingua PML2, (ii) using this provenance information, as well as supplemental log data, building an abstracted workflow representation known as a RITE network (capable of storing workflow state Relationships, Identifiers, Types, and Explanations), and (iii) visualizing the workflow log by displaying its provenance information as a directed acyclic graph and presenting supplemental explanations for individual workflow states and relationships. To demonstrate these techniques, we describe the design of a workflow explainer for the Generalized Integrated Learning Architecture (GILA) – a multi-agent platform designed to use multiple learners to solve problems such as resolving airspace allocation conflicts. We also comment on how our approach can be generalized to explain other complex workflow systems.
