|  | |||||||||
Skill Learning Using A Bottom-Up Hybrid Model
Ron Sun, Edward Merrill, Todd Peterson
The University of Alabama
Tuscaloosa, AL 35487
Abstract
This paper presents a skill learning model Clarion. Different from existing models of mostly high-level skill learning that use a top-down approach (that is, turning declarative knowledge into procedural knowledge), we adopt a bottom-up approach toward low-level skill learning, where procedural knowledge develops first and declarative knowledge develops from it. Clarion which follows this approach is formed by integrating connectionist, reinforcement, and symbolic learning methods to perform on-line learning. We compare the model with human data in a minefield navigation task. A match between the model and human data is observed in several comparisons.
1 Introduction
Skills vary in complexity and the degree of cognitive involvement. They range from simple motor movements and other routine tasks in everyday activities to highlevel intellectual skills. We want to study lower-level" cognitive skills, which have not received sufficient research attention. One type of task that exemplifies what we call low-level cognitive skill is reactive sequential decision making (Sun and Peterson 1995). It involves an agent selecting and performing a sequence of actions to accomplish an objective on the basis of moment-tomoment information (hence the term reactive"). An example of this kind of task is the minefield navigation task developed at The Naval Research Lab (see Gordon et al. 1994). This kind of task setting appears to tap into real-world skills associated with decision making under conditions of time pressure and limited information. Thus, the results we obtain from human experiments will likely be transferable to real-world skill learning situations. Yet this kind of task is suitable for computational modeling given the recent development of machine learning techniques (Sun et al 1996, Watkins 1989).
The distinction between procedural knowledge and declarative knowledge has been made in many theories of learning and cognition (for example, Anderson 1982, 1993, Keil 1989, Damasio et al. 1994, and Sun 1995). It is believed that both procedural and declarative knowl-
obstacles
agent
target
Figure 1: Navigating Through Mines
edge are essential to cognitive agents in complex environments. Anderson (1982) originally proposed the distinction based on data from a variety of skill learning studies, ranging from arithmetic to geometric theorem proving, to account for changes resulting from extensive practice. Similar distinctions have been made by other researchers based on different sets of data, in the areas of skill learning, concept formation, and verbal informal reasoning (e.g., Fitts and Posner, 1967; Keil, 1989; Sun, 1995).
Most of the work in skill learning that makes the declarative/procedural distinction assumes a top-down approach; that is, learners first acquire a great deal of explicit declarative knowledge in a domain and then through practice, turn this knowledge into a procedural form (proceduralization"), which leads to skilled performance. However, these models were not developed to account for skill learning in the absence of, or independent from, prexisting explicit domain knowledge. Several lines of research demonstrate that individuals can learn to perform complex skills without first obtaining a large amount of explicit declarative knowledge (e.g., Berry and Broadbent 1988, Stanley et al 1989, Lewicki et al 1992, Willingham et al 1992, Reber 1989, KarmiloffSmith 1986, Schacter 1987, and Schraagen 1993). In research on implicit learning, Berry and Broadbent (1988), Willingham et al (1992), and Reber (1989) expressly demonstrate a dissociation between explicit knowledge and skilled performance in a variety of tasks including dynamic decision tasks (Berry and Broadbent 1988), artificial grammar learning tasks (Reber 1989), and serial reaction tasks (Willingham et al 1992). Berry and Broadbent (1988) argue that the psychological data in dynamic decision tasks are not consistent with exclusively top-down learning models, because subjects can