Research On Knowledge Representation And Learning For Service Robots

Knowledge representation and reasoning have always been one of the research hotspots in the field of AI,and have been widely used in many fields such as search,recommendation,question answering system,etc..Service robots as a platform which integrates many AI techniques,using knowledge representation and reasoning to en-hance its cognitive ability is also a research topic.Therefore,we studied and imple-mented related applications on the Kejia and Jiajia robot.Besides,the traditional sym-bolic knowledge representation method is hard to cope with the ever-increasing knowl-edge graph.Therefore,the knowledge representation learning has gain more and more attention.This paper studies the problems of the negative sampling algorithm in the knowledge representation learning.The main work and innovations of this paper are as follows:The core of the cognitive ability of home service robots is knowledge representation and reasoning.In the KeJia service robot,the knowledge is stored by the world model and the task planning module.The world model stores the environmental information by the simplified Ontology method.This representation is weak and difficult to maintain.We developed a knowledge graph for KeJia based on RDF schema,which enhances the ability to express knowledge and preserves the dynamic changes of environmental information.The main function of Jiajia service robot is man-machine interaction.In this paper we built a tourism knowledge graph of Anhui Province,and developd a travel itinerary planning function based on this.Since the knowledge representation learning model needs positive samples and negative samples to perform a discriminate training to optimize the knowledge repre-sentation embeddings,most models use negative sampling to construct negative samples for training.But the negative samples generated by this method have low confidence,the model will converge quickly and stop optimizing.In this paper,an adversarial negative sample generation algorithm was proposed.By using information from positive sam-ples,the generator generates negative samples for discriminators to perform a adversar-ial training.In this paper,five different target models are implemented as discriminator to provide rewards for the generator,and the knowledge representation embeddings in the discriminator is used to perform link prediction and triple classification on three data sets.The experimental results show that this method can effectively improve the performance of the knowledge representation learning model.