Validating instance-based learning mechanisms outside of ACT-R

Instance-based learning theory (IBLT) has explained human decision-making in several decision tasks. IBLT works by retrieving past experiences (i.e., instances) using a subset of cognitive mechanisms from a popular cognitive architecture, ACT-R. Until recently, most IBLT models were built within the ACT-R architecture. However, due to an integrated view of cognition and ACT-R’s complexity, it is difficult to distinguish between the specific contributions of ACT-R mechanisms used in IBLT from all the other mechanisms existent in ACT-R. Also, models built within the ACT-R architecture are often difficult to explain, communicate, and reuse in other systems. This research validates the main mechanisms of IBLT when used within ACT-R and when used in isolation, outside of ACT-R. Our results show that an IBLT model performs equally well in capturing human behavior within and outside of ACT-R, demonstrating the independence of these mechanisms from any complex interaction with other mechanisms in ACT-R. We discuss the implications of our results for a modular view of cognition. © 2012 Elsevier B.V. All rights reserved. 1. Validating instance-based learning mechanisms outside 21 of ACT-R 22 Cognitive architectures are encompassing theories of cognition 23 that unify and represent a full range of human cognitive pro24 cesses from perception to action [25]. The strengths of cognitive 25 architectures are derived from a tight integration of its different 26 components, particularly as they satisfy functional constraints that 27 helps maintain the “big picture” needed to understand the human 28 mind [3]. 29 However, the goal of tightly integrating a full range of human 30 cognitive processes presents many challenges. A main challenge 31 is the complexity of representation, communication, and reuse 32 of the cognitive functions involved in modeling behavior. For 33 example, ACT-R is a hybrid cognitive architecture that derives its 34 power from the tight integration of both symbolic and subsym35 bolic mechanisms [2,3]. The symbolic mechanisms are declarative 36 knowledge represented as chunks in memory and procedural 37 knowledge represented in the form of productions or if-then rules. 38 The subsymbolic mechanisms are statistical procedures that help 39 ACT-R process the symbolic information. Although ACT-R has 40 demonstrated accuracy in representing human cognition in a large 41 ! This research is supported by the Defense Threat Reduction Agency (DTRA) Grant number: HDTRA1-09-1-0053 to Cleotilde Gonzalez and Christian Lebiere. ∗ Corresponding author at: Dynamic Decision Making Laboratory, Carnegie Mellon University, Pittsburgh, PA 15213, United States. E-mail address: [email protected] (C. Gonzalez). diversity of tasks, developing cognitive models of human behavior 42 in ACT-R has become increasingly difficult. Model development in 43 ACT-R demands not only cognitive knowledge of human behavior, 44 but also technical expertise in the architecture and a programming 45 language (e.g., LISP [5]). Thus, some of the cognitive architectures’ 46 capabilities can only be attained with disruptive technology that 47 has little to do with the goal of integration and understanding the 48 human mind. 49 Some remedies have been proposed to deal with the complex50 ity that results from the tight integration. A recent trend in the 51 field of Human–Computer Interaction (HCI) has shown the need 52 to simplify the ACT-R architecture by advocating cognitive tools 53 that are built upon it, but that help increase the usability of ACT54 R for non-programmers in developing cognitive models [7,18,19]. 55 Also, the more recent version of ACT-R (ACT-R 6.0) has proposed a 56 modular view of cognition that has made integration of different 57 components less tight. Now, it is possible to create and maintain 58 new specialized “modules” in ACT-R, which could be reused and 59 integrated into more complex systems [27]. 60 The modular view of cognition and the simplification of the 61 modeling process through HCI techniques are both important 62 approaches to deal with ACT-R’s complexity. This modular view 63 allows for the creation of unified theories that use a subset of ACT64 R mechanisms for a particular purpose or concrete types of tasks. 65 While preserving the power of unification and the robustness of 66 the architecture’s subsymbolic mechanisms, one may propose the 67 development of concrete cognitive theories that deal with partic68 ular mechanisms rather than the architecture’s full capabilities. 69 There are at least two examples of this approach. One is the Unified 7