Learning without limits

From Problem Solving towards a Unified Theory of Learning

Niels Taatgen

The full text of the thesis is available as pdf-file (3.1 Mb)
The university library also has the thesis available for download in individual chapters

Chapter 1: Introduction

Statement of the problem: how do people learn to solve complex problems. Newell's twenty questions paper. Recommendations: 1) use a unified theory 2) use a single complex problem and study all its aspects. To comply with recommendation 1, we'll use an architecture for cognition as a vehicle. To comply with recommendation 2, we'll study a class of problems that is interesting because of formal reasons: NP-complete problems. Background knowledge about Complexity Theory, examples of complex problems, introduction of the research methods: formal analysis, empirical research through protocol analysis and computer simulation using cognitive architectures. Proof why learning is necessary to give a full account of problem solving.

Chapter 2: Architectures for Cognition

The main research goal of this thesis is a cognitive model of how people learne to solve complex problems, justified by empirical data and a formal analysis. Cognitive models can be constructed using architectures of cognition. Architectures in general will be discussed in this chapter, as well as some examples, Soar, ACT-R, 3CAPS and EPIC. Some comparisons will be made and the choice of ACT-R will be justified. Also a number of approaches to learning will be discussed in detail, and a taxonomy of learning methods will be made.

Chapter 3: Scheduling

Description and analysis of the scheduling experiment. (At least) two conclusions will be made: people use rehearsal if the plan by heart, and people revise their strategies. Other desired conclusion: subjects develop a different viewpoint on the problem with experience.

The following files are available:

Verbal protocols of all participants in Dutch (pdf-file)

Detailed analysis of participant 2 in Dutch (pdf-file)

Chapter 4: Implicit versus Explicit Learning

Idea that implicit learning is learning through the standard mechanisms of the architecture, and explicit learning is learning through learning strategies.
- Model of the Tulving experiment
- Model of rehearsal and free-recall

The following files are available:

Model of the dissociation experiment (ACT-R 4.0)

Model of rehearsal (ACT-R 2.0)

Model of free-recall simulation 1

Model of free-recall simulation 2

Model of free-recall simulation 3

Model of free-recall simulation 4 without delay

Model of free-recall simulation 5 with delay

Chapter 5: Strategies of learning

When subjects are solving problems, their behavior can be classified into two rough categories: search and reflection. Both types of behaviour have different effect on the type of knowledge that is learned. A model based on dynamical growth is presented, that shows that alternating search and reflection is a rational thing to do. Some of the predictions are tested in an experiment in which subjects have to use grammars to contruct a certain word. Some developmental theories are discussed, since they may shed light on the how learning strategies themselves are learned. Fischer's theory, Karmiloff-Smith's theory, Siegler's Theory. ACT-R model of strategy creation and revision. Example for the labeled-beam task. Extension to reversal-shift learning and developmental theories

The following files are available:

Microsoft Excel 4.0 (MAC) file with the dynamic growth model

Model of the beam task (ACT-R 3.0)

Model of discrimination-shift learning (ACT-R 4.0)

Chapter 6: Examples versus Rules

Model of the Sugar Factory and the Fincham task.

The following files are available:

Model of Fincham experiment 1 (ACT-R 4.0)

Model of Fincham experiment 2 (ACT-R 4.0)

Model of Fincham experiment 3 (ACT-R 4.0)

Chapter 7: A model of scheduling

Model will combine the results of chapter 4, 5 and 6. (& will model some mental imagery). Predictions of the model will be compared to empirical data.

The following files are available:

First scheduling model (ACT-R 4.0)

Small model that demonstrates the influence of W on working memory span (ACT-R 4.0)

Second scheduling model (ACT-R 4.0)

Chapter 8:  Conclusions

Some general conclusions: a general theory of skill learning, some remarks on individual differences and development, and practical implications, mainly in the field of cognitive ergonomics and education. The usefulness of ACT-R for this topic will also be evaluated.