Multi-discrete genetic algorithm in hopfield neural network with weighted random k satisfiability

摘要

The existing Discrete Hopfield Neural Network with systematic Satisfiability models produced repetition of final neuron states which promotes to overfitting global minima solutions. Consequently, this has a negative impact to the neural network models, especially when handling real-life optimization problems. Thus, a non-systematic Satisfiability was formulated to counter this problem, which believed to have introduce more diversification of solutions in the network. However, minimal improvement was obtained due to lack of investigation on the impact of different distribution of negative literals in Satisfiability provides. Most existing works focused on different composition of literals per clause. To fill this gap, the study of Weighted Random k Satisfiability is formulated with randomized literals and desired ratio of negative literals in the logic. Before the logic is being trained in Discrete Hopfield Neural Network, a logic phase is introduced to optimally generate the right structure of Weighted Random k Satisfiability by using Genetic Algorithm with respect to the desired ratio of negative literals. Additionally, the training phase of the Discrete Hopfield Neural Network embedded Genetic Algorithm to find satisfied assignments of the proposed logic that corresponds to optimal synaptic weight management. From the conducted experiments of different optimization algorithms, the findings show that the Genetic Algorithm outperforms other algorithms in both the logic and training phases. The quality of the retrieved final neuron states achieved acceptable results.