Reservoir Computing-Based Model of Action Planning: Emulating Prefrontal Cortex Functions with Dynamic Synapses and Reward-Based Learning

In planning actions, it is crucial to remember the final goal and convert it into immediate objectives. This process is believed to rely on working memory (WM), facilitated by a recurrent network in the prefrontal cortex. In this study, we propose a mathematical model of action planning based on reservoir computing (RC), assuming that RC has characteristics similar to those of the prefrontal cortex and realizing WM. To enhance the physiological validity of the proposed model, we introduce reward-based learning and short-term synaptic plasticity (STP). The performance of the proposed model was evaluated using a path-planning task. Experimental results confirmed that the proposed model can generate several appropriate actions to reach the specified goal, demonstrating planning capabilities equivalent to those observed in monkeys. Additionally, we verified that STP contributes to the improvement of WM performance realized by RC, enhancing the task accomplishment ability and noise robustness of the proposed model. Finally, we quantitatively analyzed the representation of information in the brain when WM is appropriately realized by quantifying the amount of the task-relevant information.