Research Of Modulated Plastic Neural Network Based On Evolutionary Computing

Millions of years of evolution have produced a vast number of biological systems,and the intelligence was crowned until the evolution of human brain.Researchers in the field of artificial intelligence aim to abstract basic principles from our brains and apply them to guide the design of algorithms or models,which can reproduce the desired intelligent behavior.Among various subdomain,neural evolution uses evolutionary computing to explore the basic mechanism of brain learning,which aims at guiding the design of plasticity neural networks with learning ability.Instead of investigating learning rules by manual summarization,evolutionary computing can systematically and automatically explore the solution space where learning rules exist.The exploration of neural evolution in learning rules is mainly based on the Hebbian learning theory.By optimizing the Hebb learning rules or other forms of learning rules,the neural network is endowed with the ability of continuous learning.On this basis,the learning process is adjusted through neural modulation to make it more effective.However,due to the complexity of learning rules,general learning rules with biological credibility have not been explored yet.Artificial constraints are added during the design of learning rules and a relatively monotonous neural modulation method is used.Besides,studies focus on the homogeneous neural network and ignore the influence of different heterogeneous factors on the performance of the neural network.In this paper,we focus on the further exploration of general biological plausible learning rules,and carries out research from three perspectives: coding mode,modulation plasticity and heterogeneous neural network.The main research contents are as follows:1)Research on general learning rules based on Cartesian genetic programming algorithm.The successful evolution of plastic networks through Cartesian genetic programming into general learning rules with a high degree of biological credibility.In addition,the use of this encoding method can be used to obtain a more diverse biologically plausible neural network.2)Learning rules of fusion neural modulation.By designing a sub-network to provide modulation signals in the network model,the output modulation signals of the sub-network are fused with the model network in a highly nonlinear way.3)Heterogeneous neural network.From the perspective of evolutionary neural network,heterogeneous factors such as learning rules,network topology and activation function are encoded into the same chromosome,and the influence of different heterogeneous factors on the performance of neural network is discussed.