Periodic Reporting for period 4 - GEMS (Genetically Evolving Models of Science)
Reporting period: 2024-05-01 to 2025-02-28
Our society depends on scientific progress, which is reliant on the development of scientific theories and models. However, the development of scientific models suffers from two related problems: the ever-growing number of experimental results and scientists’ cognitive limitations (including cognitive biases). This multidisciplinary project (psychology, computer modelling, computer science and cognitive neuroscience) addressed these problems by developing a novel methodology for generating scientific models semi-automatically. The methodology is not specific to any particular discipline and can be applied to any science where experimental data are available.
The methodology treats models as computer programs and evolves a population of models using genetic programming. The extent to which the models fit the empirical data is used as a fitness function. The best models – potentially modified by cross-over and mutation – are selected for the next generation.
To demonstrate that the methodology is sound, can be used with complex experiments and can be generalised across sciences, four related strands of research were implemented. First, ‘Building New Tools’ developed and refined the methodology, and created techniques to understand and compare the evolved models. Second, ‘Explaining Human Data’ used the methodology to explain a wide range of data on human cognition. This was done in two steps: (a) data without learning (attention and working memory); and (b) data with learning (categorisation, implicit learning and explicit learning). Third, ‘Explaining Animal Data’ developed models to account for various aspects of animal behaviour, focusing on categorisation. Finally, ‘Explaining Neuroscience Data’ showed that the methodology is in principle able to account for data combining information about cognitive processes and brain structure.
In conclusion, evolutionary computational methods can be used to evolve scientific theories, producing valuable models of similar validity as those developed by human scientists. Of particular interest were the simulations showing that the GEMS methodology developed in this project can be successfully used with the cognitive architecture CHREST to evolve models of verbal learning with human data, and models of categorisation both with human and animal data.
The methodology treats models as computer programs and evolves a population of models using genetic programming. The extent to which the models fit the empirical data is used as a fitness function. The best models – potentially modified by cross-over and mutation – are selected for the next generation.
To demonstrate that the methodology is sound, can be used with complex experiments and can be generalised across sciences, four related strands of research were implemented. First, ‘Building New Tools’ developed and refined the methodology, and created techniques to understand and compare the evolved models. Second, ‘Explaining Human Data’ used the methodology to explain a wide range of data on human cognition. This was done in two steps: (a) data without learning (attention and working memory); and (b) data with learning (categorisation, implicit learning and explicit learning). Third, ‘Explaining Animal Data’ developed models to account for various aspects of animal behaviour, focusing on categorisation. Finally, ‘Explaining Neuroscience Data’ showed that the methodology is in principle able to account for data combining information about cognitive processes and brain structure.
In conclusion, evolutionary computational methods can be used to evolve scientific theories, producing valuable models of similar validity as those developed by human scientists. Of particular interest were the simulations showing that the GEMS methodology developed in this project can be successfully used with the cognitive architecture CHREST to evolve models of verbal learning with human data, and models of categorisation both with human and animal data.
Strand 1 (Building new tools) aimed to develop the programming environment and algorithms at the core of GEMS. One key aim was to simplify the generated models. This is because genetic programming (GP) suffers from bloat (excessive growth in program size), which reduces its ability to efficiently explore complex search spaces. Two techniques were developed. The first, “generationwide simplification”, replaces a less fit parent individual with a child of better fitness. We apply this simplification mechanism to all individuals of every k_th generation. The second technique, “pruning as an operator”, uses the same idea but now at the individual level rather than at the population level. This strand also developed feature extraction techniques to identify common sub-parts of models and re-use them in seeding later runs of evolution, which allows more efficient search for models. Methods were also developed for linking the GEMS methodology to a connectionist architecture and the cognitive architecture CHREST. The linking with CHREST turned out highly successful to explain human and animal data (see strands 2 and 3 below). We also combined genetic programming with genetic algorithms, making it possible to optimise models and numeric parameters simultaneously.
Strand 2 (Explaining human data) used the GEMS software for developing and interpreting models of (a) attention (delayed-match-to-sample task, spatial cueing and selective attention tasks), (b) working memory (serial recall and n-back tasks), (c) categorisation (Shephard-Hovland Task and the 5-4 category task), (d) implicit learning (artificial grammars and problem-gambling), (e) explicit learning (verbal learning experiments) and (f) value-based decision making. The combination of GEMS and CHREST turn out to be particularly powerful for accounting the data on categorisation and explicit learning with goodness of fit similar to or higher to that of the leading models in the field.
Strand 3 (Explaining animal data) obtained excellent models for categorisation, successfully simulating experiments with pigeons and monkeys in the Shephard-Hovland and 5-4 category tasks. This was achieved by reusing large parts of the code developed to account for categorisation with humans.
Strand 4 (Explaining neuroscience data) was limited to simulations showing that the basic idea of this strand -– evolving models that include a mapping from brain structures to cognitive functions –- is achievable with the GEMS methodology.
Strand 2 (Explaining human data) used the GEMS software for developing and interpreting models of (a) attention (delayed-match-to-sample task, spatial cueing and selective attention tasks), (b) working memory (serial recall and n-back tasks), (c) categorisation (Shephard-Hovland Task and the 5-4 category task), (d) implicit learning (artificial grammars and problem-gambling), (e) explicit learning (verbal learning experiments) and (f) value-based decision making. The combination of GEMS and CHREST turn out to be particularly powerful for accounting the data on categorisation and explicit learning with goodness of fit similar to or higher to that of the leading models in the field.
Strand 3 (Explaining animal data) obtained excellent models for categorisation, successfully simulating experiments with pigeons and monkeys in the Shephard-Hovland and 5-4 category tasks. This was achieved by reusing large parts of the code developed to account for categorisation with humans.
Strand 4 (Explaining neuroscience data) was limited to simulations showing that the basic idea of this strand -– evolving models that include a mapping from brain structures to cognitive functions –- is achievable with the GEMS methodology.
The project developed a novel methodology based on genetic programming to semi-automatically generate scientific models in psychology and neuroscience. It implemented a programming environment to generate models and developed algorithms simplifying the generated models, and showed that this environment can be linked to a cognitive architecture. The programming tools were applied to specific experiments in attention, working memory, explicit learning, implicit learning, categorisation and decision making. New models were generated for these experiments, including models that perform as well as or even better than the leading models in the field. Finally, it showed that the methodology can be used for developing models of both human and animal behaviour.