Neural network models are essential tools in understanding how behavior arises from information processing in the brain. Recent advances in computing power and neural network algorithms have made more complex models possible, increasing their explanatory power. However, it is difficult to make such models work: they have many configuration parameters that have to be set right for the model to work properly. Consequently, automated methods are needed to optimize them. This paper proposed an evolutionary approach to this problem. An Age-Layered Evolutionary Algorithm is introduced and evaluated by fitting training parameters for BiLex, a self-organizing map model of lexical access in bilinguals. The resulting configurations are highly optimized and able to generalize to previously unseen human data, showing that evolutionary optimization of complex models has the potential to play an integral role in cognitive modeling in the future.

PDF

In Proceedings of the International Conference on Cognitive Modeling, 2019.

@article{grasemann:iccm19, title={Evolutionary Optimization of Neural-Network Models of Human Behavior}, author={Uli Grasemann and Risto Miikkulainen and Claudia Peñaloza, Maria Dekhtyar and Swathi Kiran}, booktitle={Proceedings of the International Conference on Cognitive Modeling}, month={ }, url="http://nn.cs.utexas.edu/?grasemann:iccm19", year={2019} }