A comparative study of soar and the construction-integration model
Cathleen Susan Wharton, Clayton Lewis · 1994
Researchers in AI and HCI are weighing the merits of contrasting classes of cognitive architectures: those that feature highly structured information processed by rules (symbolic architectures), and those that feature less structured information processed by propagation of activation (associational architectures). To better understand the trade-offs one architecture was selected from each class: Soar representing symbolic architectures and the Construction-Integration Model for associational architectures. Within the framework provided by these architectures an HCI task was modelled. The particular task modelled was one of using an Automatic Teller Machine to get the balance of a checking account. This task is said to be situated: A user must rely on both knowledge from the environment and internally to perform this task. This dissertation reports what was learned from the comparison of these two architectures and the resulting ATM task models. The findings of this study are that both of these architectures appear to be more similar than not. Many similarities were noted in the twenty-one dimensions compared at the architectural level in addition to the similarities identified for the six psychological and modelling dimensions used to compare the resulting task models. From an architectural perspective, the two architectures are extremely similar. Some points of similarity include taskability and rationality. Some differences can also be seen, as in the two architectures' approach to learning (particularly for declarative knowledge), and their underlying computational methods. These differences, however, appear to go away when the resulting task models are considered. From a modelling perspective, two key results from this study were identified. The first is than the task drove the modelling more that the architecture. This finding appears to reflect the immense power and flexibility of these architectures, presumably mirroring the flexibility of humans. It further indicates that the choice of architecture, for modelling purposes, may be irrelevant: the choice of architecture may be dependent upon the task to be modelled. For example, for a learning intensive task, one might choose Soar. For modelling errors and attention issues, one might choose CI. The second finding was that the resulting models were highly similar. Thus, comparing the models at a structural level appears not to be sufficient. Instead, to tease apart the subtle distinctions of the models, it appears that we need a strong test based on performance distinctions such as timing, learning, and errorful behavior.