Conference on Cognitive Computational Neuroscience

Cognitive flexibility allows us to switch and learn different tasks simultaneously. But why does it help to learn tasks together, and what type of tasks can benefit from co-learning? We study these questions in a paradigm, in which we use tasks that are geometrically related. In a tree environment, subjects learn to find the best policies for different tasks. However, similar tasks sometimes lead to similar policies and sometimes require very different policies. We manipulate this congruence between tasks and policies as well as the number of co-trained tasks. N=46 subjects did our experiment on Prolific, and two reinforcement learning models were trained on their choices. We find that, the more tasks were learned together, the better was generalisation to new tasks in our model-based agent, but not in subjects or in the model- free agent. People and models were better at generalising to new tasks the more congruent these were to the training tasks. Hence, this has implications for curriculum design for humans and machines alike.