Research On Intelligent Learning Algorithm And Application Of Fuzzy Cognitive Maps

In the development of modern science,data mining and machine learning have become increaingly prominent,and widely affect various aspects of people's lives.Data mining algorithms deal with complex types of data,and there are correlations and effects on each other.These data provide a wealth of information,but also presents a challenge to the algorithms,so how to find the implicit relationship from observed data has become an important problem of data mining.Fuzzy Cognitive Maps?FCMs?are a kind of soft computing methods which combines fuzzy logic and neural netowrks to express the causal relationship between concepts.FCMs are widely used in modeling problems such as complex systems,decision analysis,and network structure inference.In general,data-driven optimization algorithms are used to train the weight matrix of FCMs,in this way,the causal relationship between concepts in FCMs can be quickly obtained without the help of expert knowledge,which can solve practical engineering problems more quickly.In this dissertation,the problem of intelligent learning of fuzzy cognitive maps was studied deeply from the aspects of large-scale optimization and sparse learning.In terms of large-scale optimization,current evolutionary computation based learning algorithms can only learn the problem of small fuzzy cognitive graph?less than 40 nodes?.In this dissertation,we proposed several strategies to speed up the efficiency of the learning process:?1?the combination of evolutionary computation and neural network;?2?distributed computing framework;?3?dynamic optimization strategy.In terms of learning network with sparse network structure,we propose several learning methods based on multi-objective evolutionary computation and intelligent computation based on regularization method,which is focus on the problem that the network density of the cognitive maps obtained by the algorithm is too large,which does not match the actual network density.In addition,we applied the proposed methods to the reconstruction of genetic regulatory networks?GRNs?,and compared fuzzy cognitive maps with deep neural networks?DNNs?.The following is the main work of this dissertation:1.A hybrid learning algorithm based on neural network and evolutionary computation is proposed and applied to the reconstruction of gene regulatory networks?GRNs?.The evolutionary algorithm is used to optimize the network structure,and the neural network is used to calculate the weight of the relationship between nodes in the network.At the same time,the local search operator is introduced into the algorithm,which improves the convergence efficiency.The performance of the proposed algorithm was validated on synthetic FCMs data with 20 to 100 nodes and benchmark GRNs reconstruction data DREAM3 and DREAM4.The comparison is also made with existing algorithms,including ACO,RCGA and NHL.The results show that the proposed algorithm can reconstruct the GRNs with good performance.2.An FCMs learning algorithm based on dynamic multi-agent evolutionary algorithm is proposed and applied to the problem of GRNs reconstruction.In this algorithm,the behavior of agents is designed according to the inherent characteristics of the problem.All agents live in the grid environment,and each agent's neighbor can dynamically change according to its own energy.The performance of the proposed algorithm is validated both synthetic FCMs data and GRNs reconstruction benchmark DREAM3 and DREAM4 with 5to 200 nodes.The comparison is also made with existing algorithms,including RCGA,D&C RCGA,ACO,BB-BC and DE.The results show that the proposed algorithm can effectively learnlarge-scale problems.3.An FCMs learning algorithm based on distributed dynamic multi-agent evolutionary algorithm is proposed and applied to the problem of GRNs reconstruction.The proposed algorithm uses the strategy of divided and conquer to divide the original N-nodes network into N independent sub-networks to solve the problem of dimensionality disaster encountered in our previous algorithm.At the same time,the use of distributed idea greatly improves the computational efficiency,making the algorithm suitable for large-scale FCMs learning problems.The algorithm performance is validated on synthetic FCMs data with 5to 500 nodes and GRNs reconstruction benchmark DREAM4.The comparison is also made with existing algorithms,including ACO_RD,RCGA,DE and dMAGA.The experimental results show that the proposed algorithm is effective for large-scale FCM learning problems.4.To deal with the problem that the network density of FCMs obtained by current learning algorithms is too large,an FCMs learning algorithm based on multi-objective evolutionary algorithm?MOEA?is proposed.First,the FCMs learning problem is modeled as a multi-objective optimization problem,and then the MOEA is used to solve the optimization problem.Finally,the Pareto fronts composed by FCMs with different densities are obtained.Decision makers can choose different network structure models according to their own needs.In the experiment,the performance of the algorithm is validated on both synthetic FCMs data and real data sets.The comparison is also made with existing algorithms,including SRCGA,RCGA,NHL,and DD-NHL.The results show that the proposed algorithm can learn FCMs with sparse network structures effectively.5.The FCMs learning algorithm based on ensemble learning and MOEA is proposed,termed as EMOEA_FCMs-GRNs.First,the MOEA is used to learn the Pareto fronts from the observed time series data,which contains models with different structures.Then,the networks with smaller simulation error are selected from the Pareto fronts,and an effective ensemble strategy is designed to integrate these selected networks into one network.At the same time,in the MOEA,we design a new objective function to make the networks in Pareto fronts have great difference in structure,which guarantee the performance of the following ensemble learning strategy.The experiments are performed on synthetic FCMs data and GRN benchmark DREAM4,which verifies the performance of EMOEA_FCMs-GRNs.6.An FCMs learning algorithm based on L_1/2 regularization method is proposed,which can learn large-scale FCMs with sparse structure.Based on the characteristics of FCMs learning problems,this algorithm extends the threshold-based IHTA algorithm to solve the nonlinear L_1/2 regularization problem,which is labeled as EIHTA-FCMs.The experimental results on synthetic FCMs data and real world data show that EIHTA-FCMs has a low computational cost and can effectively learn large-scale FCMs.The proposed algorithm is superior to the current FCMs learning algorithms such as SRCGA,RCGA,NHL and DD-NHL in terms of Out_of_Sample_Error,Model_Error and SS Mean.7.In terms of engineering applications,this dissertation applies the proposed FCMs learning algorithms to the problem of GRNs reconstruction.GRNs refers to the network of genes which interact with each other.In this dissertation,FCMs are used to model the dynamic control system of GRNs.The edges in the network represent the interaction between genes.The state value of nodes indicates the expression level of genes.The performance of these algorithms were validated on the GRNs reconstruction benchmark DREAM3 and DREAM4.The results show that these algorithms can solve the reconstruction of GRNs problem well.8.The similarity and difference between FCMs and deep neural networks?DNNs?in terms of structure,formal description and reaoning mechanism are analyzed and discussed.It is pointed out that the causal knowledge representation based on FCMs is expressed intuitively through the relationship between concept nodes.The dynamic behavior of the system is simulated by the interaction of concept nodes of the whole network.The reaoning is achieved by the recursive action of the forward concept node on the state of the backward concept node.At the same time,we also points out FCMs are DNNs with qualified conditions.