A decay-based account of learning and adaptation in complex skills.

Abstract

How do humans adapt to parametric changes in a task without having to learn a new skill from scratch? Many studies of memory and sensorimotor adaptation have proposed theories that incorporate a decay on prior events, which leads the agent to eventually forget old experiences. This study investigates if a similar decay mechanism can account for human adaptation in complex skills that require the simultaneous integration of cognitive, motor, and perceptual processes. In 2 experiments, subjects learned to play a novel racing video game while adapting to parametric changes in the physics of the game's controls. Human learning and performance were modeled using the ACT-R cognitive architecture, which has been used successfully to model learning and fluency across a wide range of skills in prior research. Anderson et al. ( 2019) introduced the Controller module, a new component of the architecture that learns the setting of control parameters for actions and allows the agent to execute the rapid and precise actions that are necessary for good performance on complex tasks. Model simulations support including a moderate time-based decay on the weight of the experiences that the Controller uses. This is implemented in the Controller module by discounting the influence of older observations which helps the agent to focus on recent experiences that better reflect the current relationship between different settings of a control parameter and the rate of payoff from using that setting.

Periodic tapping mechanisms of skill learning in a fast-paced video game.

“…These estimates correspond to reasonable guesses as to what people can and cannot do. Based on results from Seow et al (2019Seow et al ( , 2021, we implemented an exponential decay in the Controller. The corresponding decay parameter :tracker-decay-method was thus set to "exponential" to reflect this modeling design choice.…”

“…They are useful to experimental researchers in so far as they provide a framework for further understanding how seemingly separate neurocognitive processes are brought together to yield complex behavior. In terms of computational simulations of human skill learning, the ACT-R architecture has proved to successfully replicate key patterns of skill acquisition across various performance and motor measures in video games (Anderson et al, 2019, 2021; Dimov et al, 2023; Gianferrara, Betts, & Anderson, 2021; Moon & Anderson, 2013; Seow et al, 2021).…”

An Integrated Model of Collaborative Skill Acquisition: Anticipation, Control Tuning, and Role Adoption

“…To develop a collaborative task of a higher complexity, we started from the computer game Space Fortress (Mané & Donchin, 1989), which has an over 30‐year‐long history in research on skill acquisition (Donchin, 1995; Frederiksen & White, 1989). Recently, interest in this game has resurged with several studies relying on a PyGame implementation of the original game (Destefano & Gray, 2016; Gray & Lindstedt, 2017) and modifications of it (Anderson et al., 2019; Moon & Anderson, 2013; Seow, Betts, & Anderson, 2021), all with the goal of investigating individual skill acquisition.…”

“…In fact, computational skill acquisition models have already been built on top of cognitive architectures. These models have described diverse phenomena such as past tense learning in English (Taatgen & Anderson, 2002), learning to solve pyramid problems (analogical to arithmetic problems; Tenison, Fincham, & Anderson, 2016), and, importantly, learning to play Space Fortress and similar games (Anderson et al., 2019; Seow et al., 2021; Gianferrara et al., 2021). With this work, we will build on these modeling efforts.…”

Recent experience and performance during a critical in‐flight event

“…It is typically attributed to errors in memory and the reduction in the activation of a memory trace (Cecilio‐Fernandes et al, 2018). In the absence of rehearsal, productions or pattern‐action associations that form the basis of skilled performance degrade and/or are less accessible, resulting in errors and/or an increase in response latency over time (Seow et al, 2021).…”