Academic journal article The Behavior Analyst Today

Simulating a Shaping Task

Academic journal article The Behavior Analyst Today

Simulating a Shaping Task

Article excerpt

A neural network simulation of Eckerman, Hienz, Stern, & Kowlowitz' (1980) "Long Key" experiment showed a good match to the effective shaping of the location of a pigeon's keypeck. The neural network model was biologically based on In-Vitro Reinforcement (IVR) principles of learning. The behavioral function of the model was a "Win-Stay" strategy. Results indicated that control by consequences can account for a substantial proportion of the shaping reported in the original experiment. The simulation and its analysis are good examples of new techniques for neural network simulation, including semi-situational simulations and direct analysis.


Shaping is one of the most widely used techniques in applied behavior analysis. Despite this, very little experimental analysis of shaping has been reported from the laboratory. The present study continues the authors' efforts to use neural network simulations as a tool for the analysis of shaping. Specifically, we will show how a network based on the principles of In-Vitro Reinforcement (IVR--Stein, 1997; Stein, Xue, & Belluzzi, 1993; 1994; Stein & Belluzzi, 1989) accounts for experimental results of pigeons' key pecks shaped using a Long Key apparatus (Eckerman, Hienz, Stem, & Kowlowitz, 1980).

Across psychology, the use of computer simulations is becoming more and more common. Most of these simulations are not behavioral. And most of the methods and techniques developed to analyze these simulations are not appropriate to a behavioral analysis. Even in the case of behavioral simulations, it is not always the case that the phenomena simulated are of especial interest to the clinician. The present study uses a novel technique called direct analysis (Kemp & Eckerman, 2001), which allows the prediction of behavior from almost any experimental analysis. By selecting an experimental analysis of shaping, we hope that the present report illustrates how simulating a neural network on a computer can lead to analyses of behavior of value to the clinician.

Direct analysis means subscribing to the desiderata proposed by Church (1997). One way of speaking about direct analysis is to say it involves attempting to simulate an experiment, rather than an experimental effect. This means that the computer software that performs the simulation is composed of at least three separate programs (called modules), one modeling the experimental procedure and apparatus, another modeling the organism's sensorimotor systems, and the third modeling the behavior. The outputs of one module are input into the next, in a cycle. Kemp & Eckerman (2001) refer to this sort of simulation as situational. The third component module is the neural network model. The remainder of the software constitutes what Kemp & Eckerman call the In Situ testbed, used to evaluate the performance of the model with respect to the performance of the original organisms in the same experimental procedure.

Because the output of the behavioral module impacts the sensorimotor module, the level of detail at which the behavior is modeled must be very fine grained--both temporally and spatially. Not only each session, but each trial and each response, and often unmeasured responses (such as crossover responses and observing responses, etc.) as well, must be simulated individually. This is what we mean by simulating an experiment, rather than just an experimental effect.

Direct analysis is distinct from other types of quantitative analysis of behavior. Typical computer simulations merely calculate the functional relations. The inputs to traditional simulations are parameters describing the contingencies and the outputs are parameters characterizing the behavior. As such, they are abstractions away from the moment-to-moment details of an organism interacting with its environment (Skinner, 1976). In contrast, the output from a situational simulation system is a record of behavior with the same structure as the data recorded from the original experiment. …

Search by... Author
Show... All Results Primary Sources Peer-reviewed


An unknown error has occurred. Please click the button below to reload the page. If the problem persists, please try again in a little while.