Linear-dependent multi-interpretation neuro-encoded expression programming

Neuro-Encoded Expression Programming (NEEP) implements the continuous coding for the discrete solution through recurrent neural networks (RNNs), and smooths sharpness of the discrete coding. However, the insertion model generating linear coding in NEEP breaks the coherence of linear coding of RNNs, because the resulting symbols tend to be cluttered when RNNs learn the incoherent sequence relationships. Meanwhile, the redundancy phenomenon that different RNNs generate the same code results in that lots of solutions with the same performance exist in the search space, and causes the decrease for search efficiency. To address these problems, the linear-dependent multi-interpretation NEEP(LM-NEEP) is proposed in this research. LM-NEEP tackles the incoherence problem by employing a linear dependence strategy, and the multi-interpretation strategy is adopted to deal with the redundancy problem in search space. The capability of LM-NEEP is estimated on several symbolic regression problems. The experimental results display that the LM-NEEP significantly outperforms NEEP and some classical genetic programming methods.