A comparative study of two learning rules for associative memory

This paper addresses itself to a practical problem encountered in using iterative learning rules for associative memory models. The performance of a learning rule based on linear programming which overcomes this problem is compared with that of a representative iterative rule by numerical simulation. Results indicate superior performance by the linear programming rule. An algorithm for computing radii of maximal hyperspheres around patterns in the state space of a model is presented. Fractional volumes of basins of attractions are computed for the representative iterative rule as well as the linear programming rule. With the radii of maximal hyperspheres as weight factors for corresponding patterns to be stored, the linear programming rule gives rise to the maximal utilisation of the state space.