Developments in infants' search for displaced objects

In an exploratory study of infants' search for displaced objects, 13-month-olds and 21-month-olds were tested on three kinds of displacement problems—visible displacements, invisible displacements, and transpositions—as well as singlehiding problems. Two aspects of performance on the displacement problems were distinguished: (1) searching within the locations involved in the displacement rather than at a control location, and (2) selecting between the displacement locations. Both the younger and the older infants were able to select the correct location on visible displacement problems and to search within the displacement locations on invisible problems. Neither age group was able to solve the transposition problems, but the older infants did at least search within the relevant locations on those problems. Age differences appeared to be due primarily to improved skill at identifying relevant locations rather than to improvements in selecting among those locations. Other factors contributing to the age differences in performance were a decrease in response biases and an increase in skills for coping with multiple possibilities.     

A moving average model for sequenced reaction-time data

Reaction times are often collected in order to study the durations of hypothesized subprocesses. In many important cases, reaction times are collected for a sequence of tasks in which the processing for one task is a subset of the processing required for the next in the sequence. Data of this sort have been analyzed by multiple-regression methods. Here we propose a more sophisticated moving average model for such data. We show that ordinary multiple-regression methods produce estimates that are consistent, but not efficient, and that the usual R² statistic may provide a misleading impression of goodness of fit. One consequence of the suggested statistical model is that it allows the same quantity of data to be used to fit a psychological model to a larger range of tasks.     

Learning to program in LISP

We have gathered protocols of subjects in their first 30 hours of learning LISP. The processes by which subjects write LISP functions to meet problem specifications has been modeled in a simulation program called GRAPES (Goal Restricted Production System). The GRAPES system embodies the goal-restricted architecture for production systems as specified in the ACT* theory (Anderson, 1983). We compare our simulation to human protocols on a number of problems. GRAPES simulates the top-down, depth-first flow of control exhibited by subjects and produces code very similar to subject code. Special attention is given to modeling student solutions by analogy, how students learn from doing, and how failures of working memory affect the course of problem-solving. Of major concern is the process by which GRAPES compiles operators in solving one problem to facilitate the solution of later problems.     

Putting knowledge in its place: A scheme for programming parallel processing structures on the fly

This paper introduces a mechanism called CID, the Connection Information Distributor. CID extends connectionism by providing a way to program networks of simple processing elements on line, in response to processing demands. Without CID, simultaneous processing of several patterns has only been possible by prewiring multiple copies of the network needed to process one pattern at a time. With CID, programmable processing structures can be loaded with connection information stored centrally, as needed. To illustrate some of the characteristics of the scheme, a CID version of the interactive activation model of word recognition is described. The model has a single permanent representation of the connection information required for word perception, but it allows several words to be processed simultaneously in separate programmable networks. Multiword processing is not perfect, however. The model produces the same kinds of intrusion errors that human subjects make in processing brief presentations of word-pairs, such as SAND LANE (SAND is often misreported as LAND or SANE). The resource requirements of the mechanism, in terms of nodes and connections, are found to be quite moderate, primarily because networks that are programmed in response to task demands can be much smaller than networks that have knowledge of large numbers of patterns built in.     

Identifying the relevant aspects of a problem text

Forty-nine subjects judged the relevancy of sentence parts of a word problem (the Allsports problem). Patterns of subjects' judgments suggest three problem-solving heuristics: a SETS heuristic, a TIME heuristic, and a QUESTION heuristic. Presentation of the question before the problem tends to suppress SETS and TIME heuristics. A computer program (ATTEND) is presented to simulate subjects' behavior on the Allsports problem. The program is context-sensitive in that it can change a relevance judgment upon the acquisition of further information. Averaged subject judgments and ATTEND judgments agree for 87% of the items.     

Problem solving in semantically rich domains: An example from engineering thermodynamics

Recent research on human problem solving has largely focused on laboratory tasks that do not demand from the subject much prior, task-related information. This study seeks to extend the theory of human problem solving to semantically richer domains that are characteristic of professional problem solving. We discuss the behavior of a single subject solving problems in chemical engineering thermodynamics. We use as a protocol-encoding device a computer program called SAPA which also doubles as a theory of the subject's problem-solving behavior. The subject made extensive use of means-ends analysis, similar to that observed in semantically less rich domains, supplemented by recognition mechanisms for accessing information in semantic memory.     

A learning algorithm for boltzmann machines

The computational power of massively parallel networks of simple processing elements resides in the communication bandwidth provided by the hardware connections between elements. These connections can allow a significant fraction of the knowledge of the system to be applied to an instance of a problem in a very short time. One kind of computation for which massively parallel networks appear to be well suited is large constraint satisfaction searches, but to use the connections efficiently two conditions must be met: First, a search technique that is suitable for parallel networks must be found. Second, there must be some way of choosing internal representations which allow the preexisting hardware connections to be used efficiently for encoding the constraints in the domain being searched. We describe a general parallel search method, based on statistical mechanics, and we show how it leads to a general learning rule for modifying the connection strengths so as to incorporate knowledge about a task domain in an efficient way. We describe some simple examples in which the learning algorithm creates internal representations that are demonstrably the most efficient way of using the preexisting connectivity structure.