This thesis describes EM-ONE, an architecture for commonsense thinking capable of reflective reasoning about situations involving physical, social, and mental dimensions. EM-ONE uses as its knowledge base a library of commonsense narratives, each describing the physical, social, and mental activity that occurs during an interaction between several actors. EM-ONE reasons with these narratives by applying "mental critics, " procedures that debug problems that exist in the outside world or within EM-ONE itself. Mental critics draw upon commonsense narratives to suggest courses of action, methods for deliberating about the circumstances and consequences of those actions, and—when things go wrong—ways to reflect upon and debug the activity of previously invoked mental critics. Mental critics are arranged into six layers, the reactive, deliberative, reflective, self-reflective, self-conscious, and self-ideals layers. The selection of mental critics within these six layers is itself guided by a separate collection

The 21st century is experiencing a renewed interest in an old idea within artificial intelligence that goes to the heart of what it means to be both human and intelligent. This idea is that much can be gained by thinking about one's own thinking. Traditionally within cognitive science and artificial intelligence, thinking or reasoning has been cast as a decision cycle within an actionperception loop similar to that shown in Figure 1. An intelligent agent perceives some stimuli from the environment and behaves rationally to achieve its goals by selecting some action from its set of competencies. The result of these actions at the ground level is subsequently perceived at the object level and the cycle continues. Metareasoning is the process of reasoning about this reasoning cycle. It consists of both the meta-level control of computational activities and the introspective monitoring of reasoning (see Figure 2). This cyclical arrangement represents a higher-level reflection of the standard action-perception cycle, and as such, it represents the perception of reasoning and its control.

reasoning, reasoning about action and change This paper investigates the use of commonsense reasoning to understand texts involving stereotypical activities or scripts. We present a system that understands news stories involving four terrorism scripts. The system (1) builds a commonsense reasoning problem given an information extraction template representing a terrorist incident, and (2) uses commonsense reasoning and a commonsense knowledge base to build a model of the terrorist incident. The reasoning problem, commonsense knowledge base, and model are expressed in the classical logic event calculus. The system was developed using the MUC3 and MUC4 development data set. We present the results of running the system on the MUC3 and MUC4 test data sets, using manually generated answer key templates and templates generated automatically by two MUC4 information extraction systems. We present a detailed analysis of the models produced by the system given automatically generated templates. We present methods for answering questions based on the models produced by our system. We assess the portability of the system by extending it to handle 10 scripts frequent in Project Gutenberg American literature texts. 1

In open world applications a number of machine-learning techniques may potentially apply to a given learning situation. The research presented here illustrates the complexity involved in automatically choosing an appropriate technique in a multistrategy learning system. It also constitutes a step toward a general computational solution to the learning-strategy selection problem. The approach is to treat learning-strategy selection as a separate planning problem with its own set of goals, as is the case with ordinary problem-solvers. Therefore, the management and pursuit of these learning goals becomes a central issue in learning, similar to the goal-management problems associated with traditional planning systems. This paper explores some issues, problems, and possible solutions in such a framework. Examples are presented from a multistrategy learning system called Meta-AQUA. 1 Introduction As machine learning research addresses more sophisticated task domains, the learning issues inv...

Solving problems in many real-world domains requires integrating knowledge from several past experiences. This integration requires the dynamic retrieval of multiple experiences and the extraction of their relevant subparts. Our solution is the Multi-Plan Adaptor (MPA), a method for merging partial-order plans in the context of case-based least-commitment planning. MPA provides this ability by extracting an intermediate goal statement from a partial plan, clipping a stored plan to the intermediate goal statement, and then splicing the clipping into the original partial plan. MPA is implemented in the NICOLE multistrategy reasoning system, where it is paired with MOORE, an asynchronous, resource-bounded memory module. MOORE initially retrieves its current "best guess" but continues search, spontaneously returning a better retrieval as soon as it is found. 1. Introduction Taking advantage of past experiences is the foundation of case-based reasoning. When confronted with a problem, a c...

To construct a perpetual self-aware cognitive agent that can continuously operate with independence, an introspective machine must be produced. To assemble such an agent, it is necessary to perform a full integration of cognition (planning, understanding, and learning) and metacognition (control and monitoring of cognition) with intelligent behaviors. The failure to do this completely is why similar, more limited efforts have not succeeded in the past. I outline some key computational requirements of metacognition by describing a multistrategy learning system called Meta-AQUA and then discuss an integration of Meta-AQUA with a nonlinear state-space planning agent. I show how the resultant system, INTRO, can independently generate its own goals, and I relate this work to the general issue of self-awareness by machine.

Humans learn from their mistakes. When things go badly, we notice that something is amiss, figure out what went wrong and why, and attempt to repair the problem. Artificial systems depend on their human designers to program in responses to every eventuality and therefore typically don’t even notice when things go wrong, following their programming over the proverbial, and in some cases literal, cliff. This article describes our past and current work on the metacognitive loop, a domain-general approach to giving artificial systems the ability to notice, assess, and repair problems. The

This paper describes two forms of information visualization for mixed-initiative systems associated with team collaboration and begins to discuss how plans might be formulated to achieve the visualizations. Common understanding visualization is concerned with visualizing the information a team employs, whereas visual collaboration is concerned with visualizing the ongoing, incremental information collection, the credibility and origins of that information, and the dynamic interpersonal relationships of the team itself. The first is the more "classic" form of visualization where data and information is collected, analyzed, abstracted, and tailored for display to the user. We are concerned not only with visualization for the single user, but also with visualizing the relationship the information holds in regard to the entire team. At the level of the individual user, a mixed-initiative system must consider how to tailor the appropriate information given the user's skill,...

. This research examines the metaphor of goal-driven planning as a tool for performing the integration of multiple learning algorithms. In case-based reasoning systems, several learning techniques may apply to a given situation. In a failure-driven learning environment, the problems of strategy construction are to choose and order the best set of learning algorithms or strategies that recover from a processing failure and to use those strategies to modify the system's background knowledge so that the failure will not be repeated in similar future situations. A solution to this problem is to treat learning-strategy construction as a planning problem with its own set of goals. Learning goals, as opposed to ordinary goals, specify desired states in the background knowledge of the learner, rather than desired states in the external environment of the planner. But as with traditional goalbased planners, management and pursuit of these learning goals becomes a central issue in learning. Exam...

This paper focuses upon the content and the level of granularity at which representations for the mental world should be placed in case-based explainers that employ introspective reasoning. That is, for a case-based reasoning system to represent thinking about the self, about the states and processes of reasoning, at what level of detail should one attempt to declaratively capture the contents of thought? Some claim that a mere set of two mental primitives are sufficient to represent the utterances of humans concerning verbs of thought such as "I forgot his birthday." Alternatively, many in the CBR community have built systems that record elaborate traces of reasoning, keep track of knowledge dependencies or inference, or encode much metaknowledge concerning the structure of internal rules and defaults. The position here is that a system should be able instead to capture enough details to represent causally a common set of reasoning failure symptoms. I propose a simple model of expectation-driven reasoning, derive a taxonomy of reasoning failures from the model, and present a declarative representation of the failure symptoms that have been implemented in a CBR simulation. Such representations enable a system to explain reasoning failures by mapping from symptoms of the failures to causal factors involved.