Cognitive Agent Neuron: Learning and Decision-Making Inspired by Dopaminergic Neurons

Decision-making is a critical function of the brain as an information-processing system that can be understood using mathematical models, computer simulations, and theoretical analysis. In this paper, we introduce a computational neuron model inspired by dopaminergic functionality, which we call Cognitive Agent Neurons (CANs). Dopaminergic neurons play an important role in transmitting global reward prediction errors (RPEs) across a wide array of brain regions. RPEs serve as a key signal in adaptive learning and behavior, directly impacting reinforcement learning mechanisms. We propose a rational neuron model as a neuronal agent that can communicate with its environment and improve its decision-making skills over time, inspired by dopaminergic neurons. The network of neuronal agents can cooperate to solve any task and improve their functionality with the reinforcement learning algorithm over time. So, the CAN model is an information processing system that is based on the premise that each neuron functions as an autonomous and rational agent. We examine the functionality of the CAN model based on the classification task. Due to the chemical synapse, we proposed two Dynamic (extensive-form) games with imperfect information to model the interaction of neurons in the CAN model. After training the model with approximate Q-learning, each neuron can choose the neuron that sends its signal. The results reveal that the CAN model's accuracy rate for dividing the banknote authentication dataset is 97%, its accuracy rate for dividing two classes of the notMNIST dataset is 92%, and its accuracy rate for dividing the non-linear dataset is 96%. After training the model with approximate Q-learning, the results show that the CAN model performs better in classification due to its extensive action space. After training the model with approximate Q-learning, the results show that the CAN model is effective for classification due to its extensive action space. It performs comparably to traditional machine learning methods.