The role of external memory in a complex task: Effects of device and memory restrictions on program generation

摘要

One significant finding in programming research is that code is not generated in a linear fashion - that is, in a strict first-to-last order (Davies, 1991). Typically, programmers make many deviations from linear development, leaving gaps in the emerging program to be filled in later. Hence, the final text order of the program rarely corresponds to its generative order. Green et al (1987) proposed a model to account for this finding. Their parsing/gnisrap model describes the process by which a skeletal plan is instantiated in a programming notation. This model introduces a working memory component into the analysis of coding behaviour that forces the model to use an external medium (eg the VDU screen) when program fragments are completed or when working memory is overloaded. Hence, programs are not simply built up internally and then output to an external media with a generative order that reflects the final text order of the program. Rather, programmers will frequently need to refer back to generated code in order to recreate the original plan structure. The parsing element of the model describes this process, while gnisrap describes the generative process. Given the cognitive costs that are involved in continually evaluating and modifying generated code, we require an explanation as to why skilled programmers rely so extensively on external rather than internal memory sources. Two experiments are reported here which attempt to address this issue directly. The first experiment considers the role of working memory in the determination of strategy for novice and expert programmers. The second experiment looks at the effects upon certain error forms of restricting the kinds of manipulations programmers can make within an environment.