Open-ended tasks can be decomposed into the three levels of Newell's Cognitive Band: the Unit-Task level, the Operation level, and the Deliberate-Act level. We analyzed the video game Co-op Space Fortress at these levels, reporting both the match of a cognitive model to subject behavior and the use of electroencephalogram (EEG) to track subject cognition. The Unit Task level in this game involves coordinating with a partner to kill a fortress. At this highest level of the Cognitive Band, there is a good match between subject behavior and the model. The EEG signals were also strong enough to track when Unit Tasks succeeded or failed. The intermediate Operation level in this task involves legs of flight to achieve a kill. The EEG signals associated with these operations are much weaker than the signals associated with the Unit Tasks. Still, it was possible to reconstruct subject play with much better than chance success. There were significant differences in the leg behavior of subjects and models. Model behavior did not provide a good basis for interpreting a subject's behavior at this level. At the lowest Deliberate-Act level, we observed overlapping key actions, which the model did not display. Such overlapping key actions also frustrated efforts to identify EEG signals of motor actions. We conclude that the Unit-task level is the appropriate level both for understanding open-ended tasks and for using EEG to track the performance of open-ended tasks.

This paper shows how identical skills can emerge either from instruction or discovery when both result in an understanding of the causal structure of the task domain. The paper focuses on the discovery process, extending the skill acquisition model of Anderson et al. (2019) to address learning by discovery. The discovery process involves exploring the environment and developing associations between discontinuities in the task and events that precede them. The growth of associative strength in ACT-R serves to identify potential causal connections. The model can derive operators from these discovered causal relations just as does with the instructed causal information. Subjects were given a task of learning to play a video game either with a description of the game's causal structure (Instruction) or not (Discovery). The Instruction subjects learned faster, but successful Discovery subjects caught up. After 20 3-minute games the behavior of the successful subjects in the two groups was largely indistinguishable. The play of these Discovery subjects jumped in the same discrete way as did the behavior of simulated subjects in the model. These results show how implicit processes (associative learning, control tuning) and explicit processes (causal inference, planning) can combine to produce human learning in complex environments.

A theory is presented about how instruction and experience combine to produce human fluency in a complex skill. The theory depends critically on 4 aspects of the ACT-R architecture. The first is the timing of various modules, particularly motor timing, which results in behavior that closely matches human behavior. The second is the ability to interpret declarative representations of instruction so that they lead to action. The third aspect concerns how practice converts this declarative knowledge into a procedural form so that appropriate actions can be quickly executed. The fourth component, newly added to the architecture, is a Controller module that learns the setting of control variables for actions. The overall theory is implemented in a computational model that is capable of simulating human learning. Its predictions are confirmed in a first experiment involving 2 games derived from the experimental video game Space Fortress. The second experiment tests predictions from the Controller module about lack of transfer between video games. Across the 2 experiments a single model, with the same parameter settings, is shown to simulate human learning of 3 video games. (PsycINFO Database Record (c) 2019 APA, all rights reserved).

A theory is presented about how instruction and experience combine to produce competence in a complex skill. The theory depends critically on three components of the ACT-R architecture. The first component interprets declarative representations of instruction so that they lead to action. The second component converts this knowledge into procedural form so that appropriate actions can be quickly executed. The third component, newly added to the architecture, learns the setting of control parameters for actions through a process similar to reinforcement learning.. These three components are intermingled throughout the course of skill acquisition, providing an instantiation of Fitts’ (1964) original characterization of skill acquisition as involving gradual shifts in the factor structure of skills. The overall theory is implemented in computational models that are capable of simulating human learning in different versions of the video game Space Fortress. Other than humans, these models are the only agents, natural or artificial, capable of learning to play Space Fortress.

In an fMRI study, participants were trained to play a complex video game. They were scanned early and then again after substantial practice. While better players showed greater activation in one region (right dorsal striatum) their relative skill was better diagnosed by considering the sequential structure of whole brain activation. Using a cognitive model that played this game, we extracted a characterization of the mental states that are involved in playing a game and the statistical structure of the transitions among these states. There was a strong correspondence between this measure of sequential structure and the skill of different players. Using multi-voxel pattern analysis, it was possible to recognize, with relatively high accuracy, the cognitive states participants were in during particular scans. We used the sequential structure of these activation-recognized states to predict the skill of individual players. These findings indicate that important features about information-processing strategies can be identified from a model-based analysis of the sequential structure of brain activation.

Part- and whole-task conditions were created by manipulating the presence of certain components of the Space Fortress video game. A cognitive model was created for two-part games that could be combined into a model that performed the whole game. The model generated predictions both for behavioral patterns and activation patterns in various brain regions. The activation predictions concerned both tonic activation that was constant in these regions during performance of the game and phasic activation that occurred when there was resource competition. The model's predictions were confirmed about how tonic and phasic activation in different regions would vary with condition. These results support the Decomposition Hypothesis that the execution of a complex task can be decomposed into a set of information-processing components and that these components combine unchanged in different task conditions. In addition, individual differences in learning gains were predicted by individual differences in phasic activation in those regions that displayed highest tonic activity. This individual difference pattern suggests that the rate of learning of a complex skill is determined by capacity limits.

Using a Cognitive Model to Provide Instruction for a Dynamic Task Jungaa Moon (jungaam@andrew.cmu.edu) Department of Psychology Pittsburgh, PA 15213 USA Dan Bothell (db30@andrew.cmu.edu) Department of Psychology Pittsburgh, PA 15213 USA John R. Anderson (ja+@cmu.edu) Department of Psychology Pittsburgh, PA 15213 USA Abstract The current study used the Space Fortress game (Donchin, 1989) to study the effects of training and instruction in acquisition of complex skills. The game requires flexible coordination of perceptual, cognitive and motor components in a dynamically changing environment. We examined whether effective instruction can be developed for such a task in the same way that instruction is developed for academic tasks. Instruction was developed for certain aspects of the game based on a set of explicit procedural rules in an ACT-R model that plays the game. Participants who were given these instructions were significantly better at handling those aspects of the game that the instructions targeted. The results indicate that it is possible to perform a task analysis of a dynamic task, develop explicit instructions from the analysis, and improve target skills. The results further provide implications for designing training and instructional systems for dynamic skill acquisition. Keywords: skill acquisition; Space Fortress game; ACT-R cognitive architecture. Introduction There has been a considerable history of taking cognitive models for the performance of various academic tasks and building successful instructional programs based on them (Anderson, Corbett, Koedinger, & Pelletier, 1995; Ritter, Anderson, Koedinger, & Corbett, 2007). Much of this work has used computer-based instructional systems where instruction is potentially available after each step of the task. The evidence is sparser for similar success in non-academic, time-pressured tasks. One challenge in providing instructions in such tasks is that processing instruction often interferes performing the task. In a study by Fu and his colleagues (Fu, Bothell, Douglass, Haimson, Sohn, & Anderson, 2006), participants were provided with real-time auditory instructions on an Anti-Air Warfare Coordinator (AAWC – see also Zachary, Cannon-Bowers, Bilazarian, Krecker, Lardieri, & Burns, 1999) task, based on a cognitive model of the task. This resulted in better decisions but slower performance and so no net improvement. It was speculated that this was because of interference in simultaneously processing instruction and performing the task. In this research we investigated whether instruction, based on a cognitive model, but given prior to the performance of a task, would improve performance of the task. We chose to pursue this issue within the context of the Space Fortress game, a computer-based video game. The Space Fortress game (Donchin, 1989) was developed for the learning strategy program initiated by DARPA to investigate the effectiveness of various learning strategies in complex tasks. The underlying assumption of the program was that there are learning strategies that make practice on complex tasks more efficient. Since then the game has been used in a number of skill acquisition studies to compare the effects of various training and instructional strategies on improving performance, minimizing performance decrements under dual-task conditions or facilitating the transfer of skills to a novel task (Fabiani, Buckley, Gratton, Coles, & Donchin, 1989; Frederiksen & White, 1989; Gopher, Weil, & Bareket, 1994; Ioerger, Sims, Volz, Workman, & Shebilske, 2003; Mane, Adams, & Donchin, 1989; Newell, Carlton, Fisher, & Rutter, 1989; Whetzel, Arthur, & Volz, 2008). We have developed a cognitive model capable of performing the game and closely matching human performance (Bothell, 2010) in a modern version of the Space Fortress game developed by Destefano (2010). Perhaps because of a change from joystick navigation to key-based navigation common in modern video games, it turns out that the navigation strategy adopted by experts and incorporated in our model (as well as a model by Destefano, 2010) is different than that the optimal strategy reported by Frederiksen and White (1989). We will explore the effectiveness of off-line instruction based on our cognitive model of this navigation strategy. The Space Fortress Game The main goal of the Space Fortress game (Figure 1) is to maximize the total scores by navigating a ship to destroy a fortress multiple times and protecting the ownship from the fortress and mines. The player navigates the ship in the

ACT-R (Adaptive Control of Thought - Rational) is a theory and computational model of human cognitive architecture. It consists of a set of modules with their own buffers, each devoted to processing a different kind of information. A production rule in the core production system can be fired based on the chunks in these buffers and then it changes the chunks in the buffer of the related modules or the state of the related modules, which may leads to fire a new production rule and so on to generate the cognitive behavior. It has successfully predicted and explained a broad range of cognitive psychological phenomena and found applications in the human-computer interface and other areas (see http://act-r.psy.cmu.edu) and may have potential applications in Web intelligence. In recent years, a series of fMRI experiments have been performed to explore the neural basis of cognitive architecture and to build a two-way bridge between the information processing model and fMRI. The patterns of the activations of brain areas corresponding to the buffers of the major modules in ACT-R were highly consistent across these experiments; and ACT-R has successfully predicted the Blood Oxygenation Level-Depend (BOLD) effect in these regions. The approach of ACT-R meets fMRI may shed light on the research of Web Intelligence (WI) meets Brain Informatics (BI).

This article describes the development of a real-time model-based training system that provides adaptive ''over-the-shoulder'' (OTS) instructions to trainees as they learn to perform an Anti-Air Warfare Coordinator (AAWC) task. The long-term goal is to develop a system that will provide real-time instructional materials based on learners' actions, so that eventually the initial set of instructions on a task can be strengthened, complemented, or overridden at different stages of training. The training system is based on the ACT-R architecture, which serves as the theoretical background for the cognitive model that monitors the learning process of the trainee. An experiment was designed to study the impact of OTS instructions on learning. Results showed that while OTS instructions facilitated short-term learning, (a) they took time away from the processing of current information, (b) their effects tended to decay rapidly in initial stages of training, and (c) their effects on training diminished when the OTS instructions were proceduralized in later stages of training. A cognitive model that learned from both the upfront and OTS instructions was created and provided good fits to the learning and performance data collected from human participants. Our results suggest that to fully capture the symbiotic performance between humans and intelligent training systems, it is important to closely monitor the learning process of the trainee so that instructional interventions can be delivered effectively at different stages of training. We proposed that such a flexible system can be developed based on an adaptive cognitive model that provides real-time predictions on learning and performance.

We reduced time to detect target symbols in mock radar screens by adding perceptual boundaries that partitioned displays in accordance with task instructions. Targets appeared among distractor symbols either close to or far from the display center, and participants were instructed to find the target closest to the center. Search time increased with both number of distractors and distance of target from center. However, when close and far regions were delineated by a centrally presented "range ring," the distractor effect was substantially reduced. In addition, eye movement patterns more closely resembled a task-efficient spiral when displays contained a range ring. Results suggest that the addition of perceptual boundaries to visual displays can help to guide search in accordance with task-directed constraints. Actual or potential applications of this research include the incorporation of perceptual boundaries into display designs in order to encourage task-efficient scanpaths (as identified via task analysis and/or empirical testing).

Game playing seems to satisfy a basic craving of human cognition by exercising its fundamental abilities in a competitive setting. Therefore, it provides an excellent benchmark to study and evaluate cognitive models in tractable yet naturalistic settings that are simple and formal yet reproduce much of the complexity of real life. Poker is probably the most widely played card game, with endless variations played by millions of adherents from casual players gambling pennies to professionals competing in million-dollar tournaments. Unlike other games that emphasize one particular aspect of cognition, poker involves a broad range of cognitive activities, including:

We present a computational model of human performance on the Sustained Attention to Response Task, a computer-based task in which people must withhold responses to infrequent and unpredictable stimuli during a period of rapid and rhythmic responding to frequent stimuli. The model, formulated within the ACT-R cognitive architecture, accounts for human performance in terms of two competing strategies and the dynamic modification of priorities given competing task demands to minimise both response time and error. The model suggests that such strategic factors may be responsible for the observed speed-accuracy trade-off rather than the alternative proposal based on sustained attention.

We extended the work of Anderson et al. (in press) by modeling how people learn from “over-the-shoulder” instructions – instructions given immediately after actions were executed – while participants were working on a AntiAir Warfare Coordinator (AAWC) task. Specifically, we modeled the incremental top-down influence of instructions on the visual search process. We constructed a model that first responded to and converted the over-the-shoulder instructions into declarative memory chunks. These declarative chunks of instructions were strengthened with repeated exposures to these instructions. The model then incrementally learned to improve the selection of next track as the strengths of these declarative chunks of instructions increased, and performance declined as the instructions decayed over time. The model was able to fit the data well, suggesting that the model was able to capture the effects of the instructions on learning and performance in this dynamic task.

Adaptive control of thought-rational (ACT-R; J. R. Anderson & C. Lebiere, 1998) has evolved into a theory that consists of multiple modules but also explains how these modules are integrated to produce coherent cognition. The perceptual-motor modules, the goal module, and the declarative memory module are presented as examples of specialized systems in ACT-R. These modules are associated with distinct cortical regions. These modules place chunks in buffers where they can be detected by a production system that responds to patterns of information in the buffers. At any point in time, a single production rule is selected to respond to the current pattern. Subsymbolic processes serve to guide the selection of rules to fire as well as the internal operations of some modules. Much of learning involves tuning of these subsymbolic processes. A number of simple and complex empirical examples are described to illustrate how these modules function singly and in concert.

This research investigated whether eye movements are informative about retrieval processes. Participants learned facts about persons and locations, and the number of facts (fan) learned about each person and location was manipulated. During a subsequent recognition test, participants made more gazes to high-fan facts than to low-fan facts, and gazes to high-fan facts had a longer duration than gazes to low-fan facts. However, there was no relation between the order in which items were fixated and the relative effect of person or location fan. The effect of person and location fan on gaze duration also did not differ with whether it was the person or location being fixated. A model assuming that the process of retrieval is independent of eye movements was successfully fit to the data on the distribution of gaze durations. According to this model, the effect of fan on number of gazes and gaze duration is an artifact of the longer retrieval times for high-fan facts.

This research investigated whether eye movements are informative about retrieval processes. Participants learned facts about persons and locations, and the number of facts (fan) learned about each person and location was manipulated. During a subsequent recognition test, participants made more gazes to high-fan facts than to low-fan facts, and gazes to high-fan facts had a longer duration than gazes to low-fan facts. However, there was no relation between the order in which items were fixated and the relative effect of person or location fan. The effect of person and location fan on gaze duration also did not differ with whether it was the person or location being fixated. A model assuming that the process of retrieval is independent of eye movements was successfully fit to the data on the distribution of gaze durations. According to this model, the effect of fan on number of gazes and gaze duration is an artifact of the longer retrieval times for high-fan facts.

We present a human performance model of operator control of a simplified air traffic control task developed under the Agent-based Modeling and Behavior Representation program (AMBR). The model was implemented using the ACT-R architecture of cognition. Using a well-developed cognitive architecture provided a number of benefits: 1. The reuse of common design patterns, such as unit task decomposition and retrieval/computation dichotomy, greatly simplified and accelerated model development. 2. The model inherited parameter values from previous models and required no fine-tuning of parameters. 3. The architectural learning mechanisms provided automatic learning of situations. The resulting model was quite simple, consisting of only five declarative chunks of information and 36 production rules, while providing a highly accurate model of human performance. The model matched a wide range of performance measures, including amount and type of errors, response latency and choice percentages. Performance variability is a fundamental aspect of human behavior in complex task. The model accounted not only for the average but also for the distribution of performance through the fundamentally stochastic nature of the architecture, amplified by the interaction with the dynamic simulation environment. Although ACT-R is a goal-directed architecture, the model provided a straightforward account of multi-tasking behavior through the use of interruptions triggered by the onset of messages. The model also provided a theory of cognitive workload based on architectural primitives. The port of the model to the High Level Architecture (HLA) was relatively simple and straightforward. Providing a stochastic model of behavior, however, put strong demands on the efficiency of the simulation. The implications of such demands for HLA-enabled human performance models are discussed.