Peter C. R. Lane’s research while affiliated with University of Hertfordshire and other places

A common goal in cognitive science involves explaining/predicting human performance in experimental settings. This study proposes a single GEMS computational scientific discovery framework that automatically generates multiple models for verbal learning simulations. GEMS achieves this by combining simple and complex cognitive mechanisms with genetic programming. This approach evolves populations of interpretable cognitive agents, with each agent learning by chunking and incorporating long-term memory (LTM) and short-term memory (STM) stores, as well as attention and perceptual mechanisms. The models simulate two different verbal learning tasks: the first investigates the effect of prior knowledge on the learning rate of stimulus-response (S-R) pairs and the second examines how backward recall is affected by the similarity of the stimuli. The models produced by GEMS are compared to both human data and EPAM-a different verbal learning model that utilises hand-crafted task-specific strategies. The models automatically evolved by GEMS produced good fit to the human data in both studies, improving on EPAM's measures of fit by almost a factor of three on some of the pattern recall conditions. These findings offer further support to the mechanisms proposed by chunking theory (Simon, 1974), connect them to the evolutionary approach, and make further inroads towards a Unified Theory of Cognition (Newell, 1990).

A fundamental issue in cognitive science concerns the interaction of the cognitive "how" operations, the genetic/memetic "why" processes, and by what means this interaction results in constrained variability and individual differences. This study proposes a single GEVL model that combines complex cognitive mechanisms with a genetic programming approach. The model evolves populations of cognitive agents, with each agent learning by chunking and incorporating LTM and STM stores, as well as attention. The model simulates two different verbal learning tasks: one that investigates the effect of stimulus-response (S-R) similarity on the learning rate; and the other, that examines how the learning time is affected by the change in stimuli presentation times. GEVL's results are compared to both human data and EPAM-a different verbal learning model that utilises hand-crafted task-specific strategies. The semi-automatically evolved GEVL strategies produced good fit to the human data in both studies, improving on EPAM's scores by as much as factor of two on some of the pattern similarity conditions. These findings offer further support to the mechanisms proposed by chunking theory, connect them to the evolutionary approach, and make further inroads towards a Unified Theory of Cognition (Newell, 1990).

How can we infer the strategies that human participants adopt to carry out a task? One possibility, which we present and discuss here, is to develop a large number of strategies that participants could have adopted, given a cognitive architecture and a set of possible operations. Subsequently, the (often many) strategies that best explain a dataset of interest are highlighted. To generate and select candidate strategies, we use genetic programming, a heuristic search method inspired by evolutionary principles. Specifically, combinations of cognitive operators are evolved and their performance compared against human participants’ performance on a specific task. We apply this methodology to a typical decision-making task, in which human participants were asked to select the brighter of two stimuli. We discover several understandable, psychologically-plausible strategies that offer explanations of participants’ performance. The strengths, applications and challenges of this methodology are discussed.

Genetic programming (GP), a widely used Evolutionary Computing technique, suffers from bloat -- the problem of excessive growth in individuals' sizes. As a result, its ability to efficiently explore complex search spaces reduces. The resulting solutions are less robust and generalisable. Moreover, it is difficult to understand and explain models which contain bloat. This phenomenon is well researched, primarily from the angle of controlling bloat: instead, our focus in this paper is to review the literature from an explainability point of view, by looking at how simplification can make GP models more explainable by reducing their sizes. Simplification is a code editing technique whose primary purpose is to make GP models more explainable. However, it can offer bloat control as an additional benefit when implemented and applied with caution. Researchers have proposed several simplification techniques and adopted various strategies to implement them. We organise the literature along multiple axes to identify the relative strengths and weaknesses of simplification techniques and to identify emerging trends and areas for future exploration. We highlight design and integration challenges and propose several avenues for research. One of them is to consider simplification as a standalone operator, rather than an extension of the standard crossover or mutation operators. Its role is then more clearly complementary to other GP operators, and it can be integrated as an optional feature into an existing GP setup. Another proposed avenue is to explore the lack of utilisation of complexity measures in simplification. So far, size is the most discussed measure, with only two pieces of prior work pointing out the benefits of using time as a measure when controlling bloat.

The performance of cloud computing depends in part on job-scheduling algorithms, but also on the connection structure. Previous work on this structure has mostly looked at fixed and static connections. However, we argue that such static structures cannot be optimal in all situations. We introduce a dynamic hierarchical connection system of sub-schedulers between the scheduler and servers, and use artificial intelligence search algorithms to optimise this structure. Due to its dynamic and flexible nature, this design enables the system to adaptively accommodate heterogeneous jobs and resources to make the most use of resources. Experimental results compare genetic algorithms and simulating annealing for optimising the structure, and demonstrate that a dynamic hierarchical structure can significantly reduce the total makespan (max processing time for given jobs) of the heterogeneous tasks allocated to heterogeneous resources, compared with a one-layer structure. This reduction is particularly pronounced when resources are scarce.KeywordsCloud computingDynamic hierarchical job scheduling structureGenetic algorithmsOptimisation

A fundamental issue in cognitive science concerns the mental processes that underlie the formation and retrieval of concepts in the short-term and long-term memory (STM and LTM respectively). This study advances Chunking Theory and its computational embodiment CHREST to propose a single model that accounts for significant aspects of concept formation in the domains of literature and music. The proposed model inherits CHREST's architecture with its integrated STM/LTM stores, while also adding a moving attention window and an "LTM chunk activation" mechanism. These additions address the overly destructive nature of primacy effect in discrimination network based architectures and expand Chunking Theory to account for learning, retrieval and categorisation of complex sequential symbolic patterns-like real-life text and written music scores. The model was trained through exposure to labelled stimuli and learned to categorise classical poets/writers and composers. The model categorised previously unseen literature pieces by Homer, Chaucer, Shakespeare, Walter Scott, Dickens and Joyce, as well as unseen sheet music scores by Bach, Mozart, Beethoven and Chopin. These findings offer further support to mechanisms proposed by Chunking Theory and expand it into the psychology of music.

Is it reasonable to talk about scientific discoveries in the social sciences? This chapter briefly reviews the status of scientific research in the social sciences and some of the arguments for and against the notion of scientific discovery in those sciences. After providing definitions of “scientific discovery” and “social sciences”, the chapter notes the large variety of epistemological views and methodologies drawn on by the social sciences. It discusses the extent to which the social sciences use precise formalisms for expressing theories. Critiques of the use and reliability of the scientific method in the social sciences are discussed. In spite of these critiques, it is argued that it is possible to speak of scientific discovery in the social sciences. The chapter ends with a preview of the book.