Goal Seeking Components for Adaptive Intelligence: An Initial Assessment.

摘要: Abstract : This report assesses the promise of a network approach to adaptive problem solving in which components themselves possess considerable power. We show that designed with attention temporal aspects reinforcement learning can acquire knowledge about feedback pathways they are embedded and use this seek their preferred inputs, thus combining pattern recognition, search, control functions. A review research shows networks having these capabilities have not been studied previously. demonstrate simple elements solve types problems beyond past. An associative memory is presented retains generalization noise resistance memories previously but does require 'teacher' provide desired associations. It conducts active, closed-loop searches for most rewarding an example whcih conducted through system's external environment internal predictive model environment. The latter system capable form latent learning.