| In this dissertation, we present a single model of animal learned avoidance and approach behavior. We started with our previously verified computational model of learned avoidance behavior and extend it to be capable of both learned avoidance and approach behaviors. My contributions in developing a new computational model are a modified generalization factor, a deficit-related signal, and an action selection mechanism. In the simulation of learned approach behavior, we trained the learning system in the manner followed in the experiment where each trial starts from the entrance of maze and ends when the animal eventually reaches the food pellet. Thus, within a trial, the learning system is allowed to turn around in the maze and backtrack to the previously visited intersection. The learning system generates a decision at each intersection that is based on a learned expectation of future reward. The model replicated animal behavior in terms of number of trials necessary to learn the path to food and number of wrong turns made per trial. Meanwhile, the model keeps the capability of learned avoidance behavior. The model implements a type of temporal-difference learning algorithm, and we discuss this model in the context of existing temporal-difference learning algorithms. We also show a signal in the model has similar temporal characteristic to nigrostriatal dopamine while a second signal has similar characteristics to the firing pattern of striatal neurons. Finally we discuss how both stimulus-response reinforcement learning and stimulus-stimulus association learning can occur in animal learning. |
