Bootstrapping is a widely employed technique in the process of building highly complex systems such as microprocessors, language compilers, and computer operating systems. It could play an even more prominent role in the creation of computation systems capable of supporting intelligent agent behaviors because of the even higher level of complexity. The prospect of a self-bootstrapping, self-improving intelligent system has motivated various fields of research in machine learning. However, a robust, generalizable methodology of machine learning is yet to be found; there are still a lot of learning behaviors that no existing learning technique can adequately account for. We believe a uniform, logic-based system such as active logic [1, 2], will be more successful in the realization of this ideal. The overall architecture that we envision is as follows: a central commonsense reasoner module attends to novel situations where the system does not already have expertise, and to its own failures; it then reasons its way to solutions or repairs, and puts these into action while at the same time causing "expert" modules to be either created or retrained so as to more quickly enact those solutions on future occasions. Thus what we propose is a kind of meld between declarative and procedural techniques where the former has great expressive power and flexibility (but is slow) and the latter is very fast but hard to adapt to new situations.

Logical approaches to reasoning about agents often rely on idealisations about belief ascription and logical omniscience which make it difficult to apply the results obtained to real agents. In this paper, we show how to ascribe beliefs and an ability to reason in an arbitrary decidable logic to an agent in a computationally grounded way. We characterise those cases in which the assumption that an agent is logically omniscient in a given logic is `harmless' in the sense that it does not lead to making incorrect predictions about the agent, and show that such an assumption is not harmless when our predictions have a temporal dimension: `now the agent believes p', and the agent requires time to derive the consequences of its beliefs. We present a family of logics for reasoning about the beliefs of an agent which is a perfect reasoner in an arbitrary decidable logic L but only derives the consequences of its beliefs after some delay #. We investigate two members of this family in detail, L# in which all the consequences are derived at the next tick of the clock, and L # # in which the agent adds at most one new belief to its set of beliefs at every tick of the clock, and show that these are sound, complete and decidable.

The kind of resource limitation that is most evident in commonsense reasoners is the passage of time while the reasoner reasons. There is not necessarily any fixed and final set of consequences with which such a reasoning agent ends up. In formalizing commonsense reasoners, then, one must be able to take into account that time is passing as the reasoner is reasoning. The reasoner can then make use of such information in subsequent deductions. Step-logic is such a formalism. It was developed in [Elgot-Drapkin, 1988] to model the ongoing process of deduction. Conclusions are drawn step-by-step. There is no "final" state of reasoning; the emphasis is on intermediate conclusions. In this paper we use step-logic to model the Three-wise-men Problem. Although others have formalized this problem, they have ignored the time aspect that is inherent in the problem: a correct assessment of the situation is made by recognizing that the reasoning process takes time and determining that the other wis...

Abstract not found

Although much effort has been expended by researchers in trying to maintain a consistent belief base in formalizing commonsense reasoning, there is some evidence that the nature of commonsense reasoning itself brings inconsistencies with it. I will outline a number of sources of such inconsistencies, and discuss why they appear unavoidable.

This essay describes a general approach to building perturbation-tolerant autonomous systems, based on the conviction that artificial agents should be able to notice when something is amiss, assess the anomaly, and guide a solution into place. This basic strategy of self-guided learning is termed the metacognitive loop; it involves the system monitoring, reasoning about, and, when necessary, altering its own decision-making components. This paper (a) argues that equipping agents with a metacognitive loop can help to overcome the brittleness problem, (b) details the metacognitive loop and its relation to our ongoing work on time-sensitive commonsense reasoning, (c) describes specific, implemented systems whose perturbation tolerance was improved by adding a metacognitive loop, and (d) outlines both short-term and long-term research agendas.

This paper presents a possible solution for the problem of integrating contextual knowledge in the WordNet database. Contextual structures are derived from three sources: (1) minimal contexts - in the form of semantic nets transformations of WordNet glosses; (2) dynamic contexts rendered by webs of lexico-semantic paths revealing textual implied information and (3) static contexts - represented by patterns of concepts and semantic links. The relevance of these structures is measured on a three-tired benchmark, comprising (a) word-sense disambiguation; (b) coreference resolution and (c) acquisition of domain patterns for information extraction. 1 The Basic Idea Recently, a new version of the WordNet lexical database [Miller, 1995] developed at Princeton has become publicly available (www.cogsci.princeton.edu/~wn). WordNet 1.6 contains 126,520 English words grouped into 91,595 synonym sets, called synsets. Words and synsets are entangled by 391,885 lexico-semantic relations, making Word...

The syntactic approach to epistemic logic avoids the logical omniscience problem by taking knowledge as primary rather than as defined in terms of possible worlds. In this study, we combine the syntactic approach with modal logic, using transition systems to model reasoning. We use two syntactic epistemic modalities: ‘knowing at least ’ a set of formulae and ‘knowing at most’ a set of formulae. We are particularly interested in models restricting the set of formulae known by an agent at a point in time to be finite. The resulting systems are investigated from the point of view of axiomatization and complexity. We show how these logics can be used to formalise non-omniscient agents who know some inference rules, and study their relationship to other systems of syntactic epistemic logics,

A BDI-based continuous-time modelling approach for intracellular dynamics is presented. It is shown how temporalised BDI-models make it possible to model intracellular biochemical processes as decision processes. By abstracting from some of the details of the biochemical pathways, the model achieves understanding in nearly intuitive terms, without losing veracity: classical intentional state properties such as Beliefs, Desires and Intentions, are founded in reality through precise biochemical relations. In an extensive example, the complex regulation of Escherichia coli vis-à-vis lactose, glucose and oxygen is simulated as a discrete-state, continuous-time temporal decision manager. Thus a bridge is introduced between two different scientific areas: the area of BDI-modelling and the area of intracellular dynamics.

Logics are powerful tools with which to reason about the behaviour of AI agents. A particular logic may provide an agent with an internal language with which it can represent and reason about the world, and act based upon its internal manipulation of this language. We can also use logics to reason about the behaviour of agents from an external perspective. For example, we could use a dynamic logic to represent the interaction between a simple deterministic automaton and its environment by introducing environment state update variables into whatever logic we use. The resulting logic may state true facts about the agent-environment pair, even if the agent does not make use of our dynamic variables. The agent may, for instance, have no internal reasoning capabilities. It may simply have its inputs hard-wired to its outputs, and so operate on no internal logic at all. However, we can still reason about the behaviour of the agent using complicated logics. We can reason about...