Cognitive architectures, and the computational process models developed to run within these architectures, are a useful tool in the living lab approach. They allow one to study and predict many behavioral and performance effects of the design of systems before, during, and after the involvement of specific human-in-the-loop experimentation and testing. Not only does one get the predicted behavior, but also traces of the processes that interacted to result in such a behavior, giving any designer or engineer the opportunity to systematically change the interactions in sociotechnical systems. Developing these models often leads to a better understanding of the specific cognitive processes involved, or at least a more robust understanding of our premises, which can propagate to other portions of the living laboratory framework.
ACT-R (Anderson, 2007), Soar (Laird, 2012), and Epic (Kieras & Meyer, 1997) are discussed in the following sections. These architectures all fill particular niches in the cognitive architecture space (particularly the functional-level architecture space) and have slightly different perspectives that have colored their development. However, all architectures have notable features that can make them useful, depending on one’s need.
ACT-R is a hybrid modular cognitive architecture that uses symbolic and subsymbolic representations to store and process information. There are two main memory systems within the architecture: a procedural memory system (composed of the procedural module, utility module, and production-compilation module) and a declarative memory system (declarative module). In many ways, the architecture has been built around these memory systems, for example, the perceptual-motor modules are more recent additions to the system (Byrne, 2001). The architecture also contains a goal module, which modulates a representation of current goals, and an imagi- nal module that is responsible for imagined representations, that are also central to information processing. Figure 11.1 gives a high-level picture of the architecture. We focus on the memory systems in this overview of ACT-R; see Byrne (2001) and Anderson (2007) for a discussion of other systems in ACT-R.
The procedural memory system provides memory in the form of state-action pairs, implemented (in the software) as if-then style rules. Procedural memories, or production rules, have subsymbolic properties, with each rule having a certain utility that can be used with reinforcement-learning (Fu & Anderson, 2006) and also procedural compilation (Taatgen & Lee, 2003); procedural compilation is a way for a model to learn a skill more quickly by compiling sequences of rules into a single rule. All rules that completely match any given state (as determined by information in buffers during a stage called conflict resolution) can be fired, but only one rule (the rule with the highest at-time utility) will actually fire on a given cycle. If partial
FIGURE 11.1 A high level picture of the ACT-R architecture, the architecture contains several functional modules that use buffers for communication with the external environment and with the procedural memory system.
FIGURE 11.2 The ACT-R conflict resolution matches existing production rules against the current buffer states and picks the rules that have the highest utility.
matching is enabled within the architecture (using a parameter, :ppm) rules that only partially match the current state can also be selected during conflict resolution. Figure 11.2 gives an idea of how the current state (as specified by perceptual-motor and central-system buffers) may cause certain rules to be selected, and fired, during conflict resolution.
In Figure 11.2, the conflict resolution cycle results in a few production rules to be selected (in this case, we assume partial matching is turned off). Also during this conflict-resolution cycle, one rule is selected and fired. After symbolic pattern matching (via the current state of the buffers), subsymbolic memory comes into play through utility values attached to each rule, the rule with the highest utility is selected and fired. A form of stochasticity can be modeled and simulated using the procedural memory noise parameter (represented as egs in the canonical, LISP version of ACT-R). Explicit in the theory of the architecture is that only one rule can fire at a time, that is, procedural memory represents a serial bottleneck on central cognition. All other modules (and corresponding buffers) can operate in parallel during a given decision cycle. Thus, the architecture is mostly parallel in its operation before and after the conflict resolution.
Procedural memory selection is dependent on activity in buffers of all of the modules (central and perceptual motor). In current ACT-R models, more focus is typically placed on central modules (i.e., the goal, imaginal, and especially the declarative modules.) Declarative memory is important in this mechanistic process as these chunks of knowledge not only hold facts about the world, but also internal representations, such as goals.
Declarative memory is used by the architecture to retrieve knowledge about the world, and indeed the model itself. The module is parallel in nature both in terms of its use of symbolic memory and subsymbolic memory. Declarative knowledge is represented on a functional level as certain chunks that can have their subsymbolic value strengthened by continuous retrieval over time, and also retrieval by associated chunks. The subsymbolic activation value of each chunk in declarative memory is represented using Equation 11.1. This means that the probability of each declarative memory is potentially affected by several factors: time since last retrieval of the initial base-level activation of that memory (B,), contextual activation of that memory either from both a deliberate memory probe, that is, trying to remember something (P), or from state information contained in buffers, that is, being primed to remember something (S,), and declarative memory noise, which decreases the probability of retrieving the correct memory (e;).
As a way to illustrate the declarative-memory process in ACT-R, consider the following set of events. You are taking an exam, perhaps a stressful exam that is adding noise to your retrieval process, making it difficult to remember the correct facts. The instructor has set up the question so that it provides certain hints; the words used provide a context for the correct answer. In an ACT-R theory, these words are perceived and then processed using central modules. This processing, which includes the holding of such information in buffers, causes an increased activation of the correct declarative facts in memory that allow you to answer the question. Both similar contexts and recency will increase the probability of your retrieving the facts to solve your problem (assuming, of course, that you have correctly learned such facts).
The procedural and declarative memory systems are typically used in concert for more complex cognitive-process models. For example, the models described by Trafton et al. (2013) use multiple modules (including both memory systems), because the tasks employed tend to be more complex and have a longer temporal scale. Those models simulate learning and behavior across weeks, months, and years. Gunzelmann et al. (2011) also use a more complex model to simulate driving, but this model has less dependency on declarative learning than the procedural memory system. Dancy (2014) used both procedural and declarative memory systems and both symbolic and subsymbolic representations in those specific systems to simulate learning and decision making during a modified version of the Iowa gambling task. Despite the power mechanisms of learning within these two memory systems, process models and intelligent agents run within the ACT-R architecture tend to primarily focus (and thus use) either declarative memory or procedural memory.
ACT-R is a hybrid modular cognitive architecture that represents cognitive processes on a functional level. ACT-R evolved as psychological theory has also evolved, making it useful for those looking to have a cognitively realistic simulation with computational-process models. Though ACT-R has two memory systems, each with symbolic and subsymbolic learning, ACT-R models/agents typically focus on one of the systems and some of the learning mechanisms. Despite this, sometimes frustrating, limitation in generality of some models, the architecture as it stands, remains a reliable and useful choice for cognitive system simulations. ACT-R’s open- source code and resources like the ACT-R website (currently http://act-r.psy.cmu. edu/) that hold a database of existing sorted related publications and models, make the architecture more approachable for non-experts.
Soar (Laird, 2012) is a cognitive architecture that is centered on the goal of providing agent functionality with cognitive plausibility; matching human data is not a primary goal of Soar. Thus, although the Soar cognitive architecture does have an underlying unified theory of cognition (Newell, 1990), it also has significant influence from goals in the AI domain (e.g., models of general intelligence). In the past Soar has relied on production rules as its sole form of long-term memory, however, more recently it has been extended to provide explicit symbolic representations for long-term declarative memory (Laird, 2008).
Traditional long-term memory structures in Soar (i.e., pre Soar 9.0) consisted solely of procedural knowledge implemented as if-then style rules. Soar models select an operator based on a condition that (usually) results in an action that retrieves information to be placed into a short-term memory buffer. The amount of information that can go into this buffer is not inherently limited by the architecture and thus must be limited explicitly by the modeler.
Short-term memory is represented in a symbolic graph structure that allows the representation of symbol properties and relations. The short-term memory structures are used by the model for checking conditions of the current state of the model’s representations within a given decision cycle. Within Soar, rules are used to propose, evaluate, and apply operators. Thus a rule is used to propose an operator, which creates a symbolic structure in the short-term memory that has an associated condition for selection.
Figure 11.3 displays the structure of typical Soar agents; these computational models are composed of problem spaces. Problem spaces are composed of operators that are condition-action pairs, these operators change the state of working memory and consequently allow the agent to move between problem spaces. Agents will typically move between subproblem spaces to accomplish a subtask to accomplish an overall goal (that will be specified with the current overarching problem space). This representation can be very useful for understanding how to solve tasks, and the problem-space operator representation allows an agent to reuse many of the
FIGURE 11.3 An high-level picture of an example Soar agent. Agents are composed of problem spaces and operators that change working memory structures and transition of the agent between problem spaces.
rules (operators) within different problem-spaces (i.e., to solve different, but perhaps related, goals).
Soar provides an architecture that straddles the line between a focus on human cognition and behavior and a focus on general agent intelligence. Though it has a foundation in human cognition, Soar is built with less of a concern for simulating and replicating human experimental data than other architectures like ACT-R. Nonetheless, Soar continues to present a cognitive architecture that is attractive for developing organized intelligent agents based on a representation of human cognition. Of particular interest may be the prospect of developing intelligent agents that process some information in an environment to enhance information processing performance by human users of a system.
EPIC is a computational cognitive architecture with theories of perceptual, motor, and cognitive processing that constrain process models (Kieras & Meyer, 1997). The EPIC architecture is composed of several information processors that are responsible for handling different types of information during any given process resulting in behavior: a cognitive processor, an auditory processor, a visual processor, an ocular motor processor, a vocal motor processor, a manual motor processor, and a tactile processor.
The cognitive processor is composed of two main components: a production rule interpreter and a working memory system. Similar to the procedural-memory system in ACT-R architecture, long-term memories in EPIC are represented as production rules, that is, if-then style rules that prescribe an action given in a particular state. A model’s state is specified by the working memory system that contains representations to give an account of all perceptual and motor processors. The central production rule system in EPIC has one important distinction from ACT-R (and indeed, Soar): the production system does not have a serial constraint, production rules can be selected and fired in parallel.
The EPIC theory constrains the bottleneck to the various working memory systems. One way to think about this in different perspective is that when completing a task on a web interface, say navigating to your desired destination page, there is no limit on how many rules may fire, as determined by the state of your processor-based working memory. Instead, there is a limit to the working memory itself, which must be used to accomplish actions throughout the environment; one only has the mental (and physical) resources to move the mouse to select one particular link on a web interface. Thus, two rules may compete for the same resources, or one of the fired rules may use the needed working-memory resources, causing a bottleneck of action.
The perceptual and motor processors are very similar to those in ACT-R (in fact, ACT-R/PM, the original extension used before the new modules were absorbed into the architecture and distributed with canonical ACT-R, was based, in part, on theory from the EPIC architecture). Perceptual processors in the architecture include an auditory processor (used to perceive sounds, both as determined due to external stimuli or internal representations placed into auditory working memory) and a visual processor that controls perceptions of visual information (as well as internal visual representations placed in visual working memory). EPIC contains three motor processors: a manual motor processor (hands), a vocal motor processor (voice), and an oculomotor processor (eyes). All processors have a general preparation phase in which resources are prepared to complete the action and an actual action phase where the desired command (as specified by some action of a production is carried out).
EPIC provides a fairly powerful perception and motor system, and takes an interesting stance on production-rule firing. The architecture itself has been used in several studies to understand and predict human visual processing (e.g., Kieras & Hornof, 2014). Nonetheless, EPIC does have drawbacks that limit its usefulness in all contexts (e.g., a little stance on how learning should occur within the architecture and no learning mechanism). This and a much smaller user community than some other architectures, has led to less architectural development. Nonetheless, the architecture can be very useful for predicting the interactions between perceptual and motor processes, and a simulated system.