Using Tactic-Based Learning (formerly Mentoring) to Accelerate Recovery of an Adaptive Learning System in a Changing Environment

Tactic-Based Learning (TBL), formerly referred to as mentoring, is a selection policy for statistical learning systems that has been initially tested with a Collective Learning Automaton that solves a simple, but representative, problem. To respond to an immature stimulus that does not yet have a high- confidence response associated with it, TBL hypothesizes that selecting a response that has been designated as useful by a different, but nonetheless well-trained stimulus, is a better strategy than selecting a random response. TBL does not use any feature analysis in search of an appropriate response. Previous results [1] show that TBL significantly accelerates learning of a static problem, especially when several stimuli share the same response, i.e., when broad domain generalization is possible. This paper shows that TBL also increases the speed of recovery when the problem changes abruptly after the learning agent has mastered the initial state of the problem.