AKF-SR: Adaptive Kalman filtering-based successor representation

To understand animals’ behavior in finding relations between similar tasks and adapting themselves to changes the tasks, it is necessary know how brain generalizes learned knowledge from a previous task unseen tasks. Recent studies neuroscience suggest that Successor Representation (SR)-based models provide adaptation goal locations or reward function faster than model-free algorithms, together with lower computational cost compared of model-based algorithms. However, not known such representation might help animals manage uncertainty their decision making. Existing methods for SR learning based on standard temporal difference (e.g., deep neural network-based algorithms) do capture about estimated SR. In order address this issue, paper presents Kalman filter-based framework, referred as Adaptive Filtering-based (AKF–SR). First, approach, which combination filter method, used within AKF–SR framework cast procedure into filtering problem benefit estimation SR, also decreases memory requirement sensitivity model’s parameters comparison An adaptive approach then applied proposed tune measurement noise covariance mapping most important affecting filter’s performance. Moreover, an active method exploits form behaviour policy leading more visits less certain values improve overall performance agent terms received rewards while interacting its environment. Experimental results three reinforcement environments illustrate efficacy over state-of-the-art frameworks cumulative reward, reliability, time cost, speed convergence function.