Tactic-based learning for collective learning systems
P. Bock, Alice J. Armstrong · 2008
Tactic-Based Learning is a new selection policy for statistical learning systems that has been tested with a Collective Learning Automaton which solves a simple, but representative problem. Current selection policies respond to immature stimuli that do not yet have high-confidence responses associated with them by selecting responses randomly. Albeit unbiased, this policy ignores any confident information already acquired for other well-trained stimuli. To exploit this confident information, Tactic-Based Learning hypothesizes that in the absence of a sufficiently confident response to a given stimulus, selecting a confident response to a different, but nonetheless well-trained stimulus is a better strategy than selecting a random response. Tactic-Based Learning does not require any feature comparison in search of an appropriate response. Preliminary results show that Tactic-Based Learning significantly accelerates learning and reduces error, especially when several stimuli share the same response, i.e. , when broad domain generalization is possible. Tactic-Based Learning reduces the use of pseudo-random number generators in the response selection process. Additionally, Tactic-Based Learning assists the recovery of learning performance when the problem evolves over time.