Snap-drift: Real-time, Performance-guided Learning

A novel approach for real-time learning and mapping of patterns using an external performance indicator is described. The learning makes use of the 'snap-drift' algorithm based on the concept of fast, convergent, minimalist learning (snap) when the overall network performance has been poor and slower, cautious learning (drift towards user request input patterns) when the performance has been good, in a non-stationary environment where new patterns are being introduced over time. Snap is based on Adaptive Resonance; and drift is based on Learning Vector Quantization (LVQ) [1], The two are combined in a semi-supervised system that shifts its learning style whenever it receives a change in performance feedback. The learning is capable of rapidly relearning and restabilising, according to changes in feedback or patterns. We have used this algorithm in the design of a modular neural network system, known as Performance-guided Adaptive Resonance Theory (P-ART) [2,3]. Simulation results show that it discovers alternative solutions in response to a significantly changed situation, in terms of the input vectors (patterns) and/or of the environment, which may require the patterns to be treated differently over time.