Study On Cognitive Development Research Based On Self-Organizing Incremental Neural Network For Robots

Since the rise of artificial intelligence,the hope that robots have the same level of intelligence as human has become an enduring issue in the field of robotics.Cognitive robots are intelligent robots that can imitate human perception and behavior and have the ability of continuous learning.At present,cognitive robots have received great attention in the frontier exploration of artificial intelligence.Different from traditional robots,cognitive robots can effectively adapt to different environments and perform varied tasks.Cognitive development ability helps robots to improve their interaction level with humans and autonomous learning ability,which promotes service robots to be truly intelligent.Excellent cognitive development models need to meet the requirements of online learning,effectively receiving new knowledge,reflecting on what has been learned and so on,so as to develop the concept of objects.However,the current cognitive development algorithms lack of autonomous feedback and means solving errors,and the similarity measurement standard is single,which is prone to classification errors.In view of the above problems,this thesis conducts an in-depth study on the online real-time cognitive development model of robots based on Self-Organizing Incremental Neural Network,which mainly includes the following:(1)In view of the defects of existing cognitive development models,which do not retain initial samples in the learning process and cannot use original data to correct the conflicts of multi-modal relational knowledge,a self-organizing incremental neural network cognitive development model combined with prototype assistance is proposed,which uses initial samples to help the network to independently correct the conflicts.The model is divided into two parts:the learning network and the prototype library.The learning network consists of three layers,namely,the sample layer,the symbol layer and the association layer,and each mode has a corresponding processing channel.The sample layer learns samples by unsupervised method and forms clustering,while the symbol layer abstracts the clustering into symbolic representation.The association layer associates the symbolic representation of each mode.The prototype library stores the original features of the samples.When the association layer finds the newly learned association relation conflicts with the previous knowledge,the input sample is compared with the central node of the corresponding category of each mode of the sample layer,and the knowledge conflict is corrected according to the comparison results.Experiments in the open environment and closed environment were carried out on fruit and vegetable dataset.The experimental results show that the proposed model can effectively learn new knowledge and make use of learned knowledge to correct errors and form object concept cognition.(2)Most of the existing cognitive development models use a single Euclidean distance to measure similarity when judging which category a sample belongs to.It is easy to make mistakes when the differences within the class are large and the differences between the classes are small.To solve this problem,a self-organizing incremental neural network cognitive development model combined with fuzzy inference is proposed,which combines incremental neural network and fuzzy inference algorithm.The model is divided into two layers.The learning network of the first layer conducts online learning of single feature of samples.In the learning process,the Euclidean distance,the activation frequency of nodes and the activity level of the local network are considered comprehensively.After fuzzy processing of the three parameters,the node with the highest similarity is selected as the winning node by fuzzy inference,which effectively solves the problem of insufficient reliability of a single measure.And when calculating the threshold value of network nodes,the method of only relying on Euclidean distance is also abandoned,the network takes the above three criteria into account.The second layer of the model is the association network,which aims to establish the correlation between different modes.The experimental results show that the algorithm has a good learning effect on both low and high dimensional data,and can effectively blur the learning process and correctly identify the categories.