On Introspection, Metacognitive Control and Augmented Data Mining Live Cycles

arXiv:0807.4417v2 [cs.AI] 15 Jan 2009 On Introspection, Metacognitive Control and Augmented Data Mining Live Cycles Daniel Sonntag German Research Center for Artificial Intelligence 66123 Saarbr¨ucken, Germany sonntag@dfki.de Abstract. We discuss metacognitive modelling as an enhancement to cognitive modelling and computing. Metacognitive control mechanisms should enable AI systems to self-reflect, reason about their actions, and to adapt to new situations. In this respect, we propose implementation details of a knowledge taxonomy and an augmented data mining life cy- cle which supports a live integration of obtained models. Keywords: Metacognitive Modelling, Data Mining 1 Introduction Cognitive computing is the development of computer techniques to emulate hu- man perception, intelligence, and problem solving. Cognitive models are equipped with artificial sensors and actuators which are integrated and embedded into physical systems or ambient intelligence environments to act in the physical world. The goal is to have cognitive capabilities and to perform cognitive con- trol (e.g., see [1]). To overcome problems in shared control (of, e.g., navigating robots [2]), direct communication (in natural language dialogue) between a hu- man participant and a technical control architecture can be employed. This could be used for mutual disambiguation of multiple sensory modalities in a learning environment. As one of the major topics of sensory-based control mech- anisms, automatic perception learning by introspection and relevance feedback could help in this disambiguation task. In order to pursue the idea of cogni- tive systems able to self-reflect, reason about their actions, and to adapt to new situations, metacognitive strategies can be employed. In this paper, we will present the core idea of a metacognitive control model of machine learning with respect to problem solving capabilities to be exemplified by improving autonomous reaction behaviour. We start by clarifying the term metacognition. Metacognition is cognition about cognition. It can, in principle, enable artificial intelligence systems to monitor and control themselves, choose goals, assess progress, and adopt new strategies for achieving goals.1 [4] associates metacognitive components with the 1 For example, students preparing for an exam judge about the relative difficulty of the learning material and use this for study strategies. The resulting reasoning task ability of a subject (or an intelligent agent in general) to orchestrate and monitor knowledge of the problem solving process; [5] argues that metacognitive abilities correlate with standard measures of intelligence; [6] talks about systems that know what they are doing. Here, we adopt the growing interest in metacognitive strategies2 for AI sys- tems to build a metacognitive model for adaptable AI systems, which involves computational models of self-representation and self-awareness. Ontologies rep- resent the knowledge groundwork for the self-representation of a system informa- tion state to be included into a metacognitive model.3 For example, McCarthy defines the term introspection as a machine having a belief about its own mental state rather than a belief about propositions concerning the world. According to this explanation of metacognition we hypothesise that researchers in adaptable AI systems should investigate in metacognition because it can help us: 1. address the difficulty to write down control management rules. Rules may not be obvious, tangible, or identifiable, or they may present an engineering overhead. 2. provide self-improvement through adaptation and customisation. 3. offer designs for never-ending learning. 4. integrate a variety of previously isolated findings: dialogue architectures, finite state strategies, information states, (un)supervised learning, stacked generalisation, reinforcement learning, interactive learning, and embedded data mining. Apart from its complexity, metacognition highlights an empirically tractable model creation and verification process. 2 Model, Introspective View and Control We use the term model in the sense given by [7]: To an observer B, an object A• is a model of an object A to the extent that B can use A• to answer questions that interest him about A. A can be the world or a specific sub-domain such as the football domain. To answer questions about the football domain, an A• has to be constructed. is a second-order reasoning process about the own learning abilities called meta- reasoning or, more generally, metacognition. 2 IBM Autonomic Computing Initiative, http://www.research.ibm.com/autonomic/, and, e.g., DARPA Information Processing Technology Office on Cognitive Systems, http://www.darpa.mil/ipto/thrust areas/thrust cs.asp. 3 [8] outlines that for intelligent behaviour, a declarative knowledge model must be created first. Examination of, e.g., own beliefs would then be possible when the beliefs are explicitly represented. McCarthy sees introspectio

…(Full text truncated)…

This content is AI-processed based on ArXiv data.