TR-66: On the Futility of Arbitrarily Increasing Memory Capabilities of Stochastic Learning Automata

Abstract

By designing a new family of expedier.t learning machines, we prove the counter-intuitive fact, that arbitraril) increasing the memory of a learning automaton does not necessarily increase its learning capability. This complements the results of Hellman and Cover [1] who proved that it is impossible to obtain an arbitrarily accurate learning machine which possesses only finite memory.