Extending the Influence of Contextual Information in ACT-R using Buffer Decay

Extending the Influence of Contextual Information in ACT-R using Buffer Decay Robert Thomson (thomsonr@andrew.cmu.edu) Stefano Bennati (sbennati@andrew.cmu.edu) Christian Lebiere (cl@cmu.edu) Department of Psychology, Carnegie Mellon University 5000 Forbes Avenue, Pittsburgh, PA, 15213, USA Abstract An Overview of the ACT-R Architecture In this paper, we describe an extension of the theory of short- term memory decay for the ACT-R cognitive architecture. By including a short-term decay for elements recently cleared from active memory, we have extended the functionality of spreading activation as a source of implicit contextual information for the model. In ACT-R models of serial memory and decision-making, contextual information has generally been modeled using either explicit markers (e.g., positional indices) or fixed-length windows of prior elements (e.g., a lag-based representation). While markers and fixed- length windows do capture some patterns of human errors, they are inflexible, are set by the modeler and not the model, and are not psychologically-plausible representations of contextual information. In conjunction with our associative learning mechanism (Thomson & Lebiere, 2013), we show how buffer decay can provide more flexible and implicit contextual information which explains refraction, positional confusion errors, and repetition facilitation and inhibition. Keywords: cognitive architectures; human memory; context ACT-R is a cognitive architecture defined as a set of modules which are integrated and coordinated through a centralized production system (see Figure 1). Modules access information from buffers, while the production system only responds to the contents of the buffers and not the internal processing of the modules themselves. The set of buffers therefore implicitly constitute the working memory of the architecture. The declarative memory and production system modules store and retrieve information that corresponds to declarative and procedural knowledge, respectively. Declarative knowledge is the kind of knowledge that a person can attend to, reflect upon, and usually articulate in some way, while procedural knowledge consists of the skills we display in our behavior, generally without conscious awareness. Introduction Over the last 50 years there has been a substantial body of literature describing how contextual information interacts with memory encoding and recall. More specifically, when the context changes between encoding and retrieval time, recall is relatively reduced compared to when retrieval occurs in the same context as encoding. Beginning with Godden and Baddeley’s (1975; 1980) seminal work on context-dependent recall in natural environments (see also Smith & Vela, 2001), research has shown both internal-state (e.g., physiological) and external-cue (e.g., environmental) dependence on recall (Eich, 1980). For instance, Godden and Baddeley found that when deep sea divers learned a list underwater, they experienced reduced list recall when recalling this list on the surface as compared to recall while underwater again. Other examples of context-dependence include mood-dependence (Eich, Macaulay, & Ryan, 1994), language-dependence (Marian & Neisser, 2000), and motivation-dependence (Delgado, Stenger, & Fiez, 2004). While context is a complex real-world phenomenon, it is much more constrained in a cognitive architecture such as ACT-R, where the information flow between a model and the environment is abstracted to a set of symbolic elements (see Anderson, Bothell, Byrne, et al., 2004). In this sense context is limited to the information available in the buffer system and the spread of activation from those items currently in working memory. Before delving more deeply into the role of contextual information in ACT-R we will present a brief overview of ACT-R and describe how context has previously been modeled in tasks involving serial memory and decision-making. Figure 1. An overview of ACT-R’s modules and their dependent buffers. Declarative knowledge in ACT-R is represented formally in terms of chunks, which corresponds to the episodic and semantic knowledge that promotes long-term coherence in behavior. Chunks have an explicit type, and consist of a set of slot-value pairs of information. Chunks are retrieved from declarative memory (DM) by an activation process. When a retrieval request is made the most active matching chunk is returned, where activation is computed as the sum of base- level activation, spreading activation, mismatch penalty and stochastic noise. Base-level activation reflects a chunk’s recency and frequency of occurrence. Activation spreads from the current focus of attention through associations among chunks in declarative memory. These associations are built up from experience, and reflect how chunks co- occur in cognitive processing. Chunks are also compared to