Analysis And Intelligent Optimization Of Network Structures For Evolutionary Games Methods

Complex networks have been widely applied to the analysis of complex systems,representing the interaction of different components.Complex networks can help analyze problems arranging from interpersonal networks to the ecosystem.The interaction of components on the complex network pushes the development of the system.This thesis focuses on the complex networks,analyzes the impact of the complex networks upon complex systems and combines the complex networks with intelligent optimization methods.To understand the emergence of cooperation,researchers analyze the impact of network structure upon evolutionary game results and try to search for structures with high cooperation level.We analyze the impact of network structures upon the evolution of cooperation strategy and optimize the network structure for evolutionary games based on intelligent optimization methods.Moreover,we also analyze how to cope with problems in the field of intelligent optimization based on complex networks.The main work in this thesis can be summarized as follows: 1.As the optimization of large networks for evolutionary games is difficult,we investigate the impact of communities on the evolution of cooperation strategy.Based on the design method proposed by Holme and Kim,we design a social network comprised by communities whose cooperation level is tunable.We analyze the correlation between the cooperation of each community and the overall cooperation level as well as the impact of communities towards the strategy evolution.The experimental results show the optimization of evolutionary game structure with communities can be realized through optimizing its communities.Moreover,the impact of a large community is larger than the coeffect brought by several small communities,even if the number of participant nodes is the same.2.We analyze the impact of diversity introduced by complex network towards cooperation,under the lack of reputation and punishment mechanisms.Referring to the social diversity,we design strategy diversity based on the complex network.We analyze the impact of strategy diversity upon the evolution of cooperation with or without preference for a longterm payoff.Experimental results show that the strategy diversity help promote cooperation,but its effect is influenced obviously by the stability of strategies under extreme conditions.When participants prefer long-term payoff,stability of strategies can be promoted and this helps strategy diversity promoting cooperation under different game parameters.3.We design an evolutionary algorithm to optimize evolutionary game structures.Optimization of complex networks is a non-deterministic polynomial(NP)hard problem.To optimize network structure for evolutionary games,it is also important to cope with problems caused by the evaluation with large variance.Traditional evolutionary algorithms rely on large resampling numbers to determine the quality of solutions and to avoid the collapse of algorithms.The multi-level evolutionary algorithm prioritizes sampling solutions with potential higher quality to reduce the number of sampling and utilizes restoration lists to adjust the evolving population to avoid the collapse of the algorithm.This helps reduce the number of sampling required to achieve a successful optimization.We test the performance of the designed algorithm in optimizing evolutionary game structures under different strategy update rules.The experimental results show the multi-level evolutionary algorithm can optimize network structures for evolutionary games under different situations.4.To promote the efficiency of algorithms when optimizing network structures for evolutionary games,we design prior experience for optimizing evolutionary game structures through analyzing the spread of strategy.Based on the microanalysis of the survival and spread of cooperation under prisoner's dilemma game,we introduce the influence transmission structure to help analyze the spread of cooperation and design a multi-level scale-free network model with high cooperation level.We summarize prior experience for optimizing network structures for evolutionary games through analyzing the correlation between cooperation and the coverage of influence,which is applied to modify the algorithm.We test the performance of the modified algorithm under different strategy update rules and compare the modified algorithm with the multi-level evolutionary algorithm.The experimental results show the advantages of the modified algorithm in optimization efficiency,which indicates the effectiveness of the prior experience.5.As available methods for customized optimization strategies can not make a balance between customization and generalization,we combine evolutionary algorithms with deep learning methods and design a deep evolutionary convolution network.We embed the optimization process of evolutionary algorithms into deep neural networks based on the commonalities between convolution and crossover operation.The deep evolutionary convolution network learns optimization strategy through updating the parameters of the model,which essentially optimizes the information flow network of evolutionary operation.As evolutionary algorithms are based on utilizing available information,the strategy learn by our model has a better generalization.We test the optimization efficiency and the generalization of the strategy learned by the model.Experimental results show the feasibility of our method in designing customized strategies for different problems.Moreover,the strategy learned has a satisfactory generalization capability.6.As the deep evolutionary convolution network cannot provide customized strategies for multiple functions within one model,we design a model that has coped with this problem based on the map between the feature of the optimization problem and the information flow network.The corresponding model extracts the feature of the optimized function based on the initial population,infers strategy for optimization based on the map between the feature space and the strategy space,and finally optimize the target function.We have analyzed the optimization efficiency of strategies given by the designed model when providing strategy for multiple functions.Moreover,we also analyze the performance of the designed model when inferring strategy for a new function and its potential value in excavating history data and providing prior experience for optimization.The experimental results show the outstanding performance of the designed model and verify its value in excavating history data and providing prior experience for optimization.7.Based on the analysis and optimization of complex networks,we design a multi-leader particle swarm algorithm with better exploration capability.The correlation network between particles and memories in conventional particle swarm optimization is easy to cause a redundant search.Multi-leader particle swarm optimization promotes the memory structure of particle swarm to provide more global and local memories for the adjustment of search patterns.Besides,it employs a multi-leader mechanism to guide the selection process of particles upon memories.These make the correlation network between particles and memories more flexible and benefit the exploration.We test the performance of the designed algorithm upon CEC-2013 and the reconstruction of the gene regulatory network,and compare it with available optimization algorithms.The experimental results show the designed algorithms have obvious advantages over available algorithms.8.To promote the stability of deep neural networks in time series prediction,we introduce the learning of complex networks into the design of deep neural network architectures.Conventional deep neural networks are hard to cope with data deviated from the training set.We integrate the learning of fuzzy cognitive maps into deep neural networks,which helps learn the complex system reproducing the observation and predict more reasonably over data deviated from the training set.In our experiment,we compare current popular deep neural networks with our design over twelve time series benchmarks and analyze the advantages of the designed model.The results show that we can promote the stability of deep neural networks in time series prediction through embedding the learning of complex networks.