Browse Books

Reviewer: Christopher K. Riesbeck

Although this book, which describes in detail a series of experiments by the authors and their colleagues on how learning transfers from one skill to another, is primarily for cognitive psychologists and cognitive scientists, it will interest many computer scientists as well. First, two of the experimental domains are computer-based, namely text editing and LISP programming (the other domain is calculus). Issues discussed include what kind of skills transfer between different text editing interfaces, what kinds of interference occur, and what kinds of mistakes novice LISP programmers make. The authors try to avoid entering religious wars about such questions as which text editor is best and whether recursion is a good thing. Second, artificial intelligence researchers will be interested in the computational model of learning, implemented in Ander s on's ACT* production system, on which the authors' experimental predictions and explanations are based. The model compiles use-independent declarative knowledge during practice into use-specific procedurally oriented production rules. When learning a new skill, the initial effects of transfer are due to overlap in declarative elements between domains, but sustained skill transfer comes from overlap in specific productions within the domain. The authors make at least two controversial claims in this book, but they defend them well. The first is that interference, or negative transfer of skills, is a minor effect that is more than compensated for by positive transfer effects. If so, then perhaps some of us should stop worrying about the debilitating effects of BASIC as a first programming language. The second, larger claim is that the ACT* model is capable of supporting quantitative predictions about human execution times and error rates. Most cognitive modeling work in artificial intelligence settles for much weaker qualitative predictions of human behavior, such as predicting that certain forms of reasoning will be “substantially” harder than others. Singley and Anderson are remarkably audacious in how far they push their model's predictions, given that (1) the procedural versus declarative distinction has been mostly abandoned in artificial intelligence as being too fuzzy, (2) they have only looked at a few cognitive skills, and (3) the quantitative predictions are very sensitive to details of how knowledge is represented. They address these criticisms in chapters 7, 8, and 9, respectively. This book is a fine example of nuts and bolts cognitive science. The focus is tight, the background discussions are thorough, the experimental designs are careful, and the statistical analyses are numerous. This is certainly not the last word on transfer, but it will be a hard act to follow.