ARENA, a rule evaluating neural assistant, has been integrated with a logic optimization system and used to determine which logic transformation out of many available should be used. Starting with no prior knowledge, the network learned to recognize the proper time to apply logic optimization rules and to select a useful rule accordingly. By selectively repeating previously learned patterns, the network was trained to choose rules which were not effective by themselves, but which were effective as a group. The network learned to reduce the training circuit's optimization evaluation value from 43 to 6. The trained network then reduced a test circuit's evaluation value from 138 to 40 using 35 operations, 10 fewer than a local-search method. The results demonstrate the suitability of using neural networks to assemble a simple sequence of actions into more complex actions toward a predefined goal.
展开▼
