A.N.A.S.A. I1 : A NOVEL, REAL-VALUED, REINFORCEMENT ALGORITHM FOR NEURAL UNIT / NETWORK. ( ANASA stands for : Adaptive Neural Algorithm of Stochastic Activation.)
Athanasios V. Vasilakos, Nikolaos Loukas, Konstantinos C. Zikidis · 1993
In this paper we present ANASA 11, a powerful algorithm for real-valued function learning, suitable for implementation as a neural network processing unit. The only information it receives from the environment, is a reinforcement, that is a number in [0,1], which indicates, in the general case stochastically, how close was the last output value to the desired one, for this particular input. The introduced innovation is that for the weight updating, apart from the 1:urrent reinforcement, the algorithm also uses the last activation and reinforcement for this input, and the average of the last (e.g. 100) reinforcements. The output is a stochastic function of the inputs and the weights, with it changing variance. The learning rate is also changing. The algorithm converges surely and smoothly, an order of magnitude faster than the best known in the literature up to now, similar algorithm, SRV [3]. Our algorithm, ANASA 11, is an improvement of the original scheme, ANASA [l].