An optimization approach to the analysis of generalized learning automata algorithms

Weak convergence methods are used to analyse generalized learning automata algorithms. The REINFORCE algorithm has been analysed. It is shown by an example that this algorithm can exhibit unbounded behaviour. A modification based on constrained optimization principles is proposed to overcome this problem. The relationship between the asymptotic behaviour of the modified algorithm and the Kuhn-Tucker points of the related constrained optimisation problem is brought out.