Research On Linguistic Reasoning And Rule Leaning Algorithms Under Uncertainty Environment

Artificial intelligence(AI)with uncertainty has been one of the hottest topics in recent years where linguistic value representation is common.Uncertain reasoning and rule learning algorithm,the important research content of AI,provide the mathematical theoretical basis for intelligence information processing and the methodology for promoting the development of AI with high reliability and high explanability.This thesis studies the approximate reasoning with linguistic value and machine learning algorithm in belief rule base under uncertain environment.The main research works can be summarized as follows:People prefer linguistic value in nature language to numerical value when they express the qualitative information under uncertainty environment.Quantitative membership function is used to handle the qualitative knowledge in fuzzy theory.It brings difficulties in representation and reasoning.Linguistic lattice implication algebra is more accordance with the characteristic of language and thinking.In this paper,the linguistic information is proceed based on the lattice order linguistic value.We extend the fuzzy modus ponens and fuzzy modus tollens onto linguistic-valued lattice implication algebra,and propose an approximate reasoning approach.The operations in lattice implication algebra are discussed for linguistic weight problem in the rules.A linguistic-real valuation function which is a positive valuation function is introduced to process incomparable linguistic ranking of the results of approximate reasoning.Illustrating examples show the effectiveness of the proposed approach which can rank the incomparable elements elastic.Linguistic terms evaluations are always used from two opposite sides at the same time.However,the relationship between the positive evidence and negative evidence is ignored in the knowledge representation and reasoning model with linguistic-valued credibility.Therefore,we propose an approximate reasoning approach with linguistic-valued intuitionistic fuzzy credibility(LV-IFC)based on linguistic-valued intuitionistic fuzzy lattice to process these linguistic knowledge.We give a knowledge representation model with LV-IFC,where the positive evidence and negative evidence can be represented and reasoning in a single rule to get the sub-conclusion.For the multi-rule,we propose a rule aggregation operator to get the final conclusion by combining all the sub-conclusions.An intuitionistic linguistic-real valuation function which is defined implying a linguistic intuitionistic fuzzy distance to rank the incomparable linguistic values if it is necessary.The contrast example shows the approach is more rational and flexible for uncertain linguistic information processing.The proposed model simplifies the linguistic knowledge representation and considers the mutual restriction relationship between the positive evidence and negative evidence.In the above approximate reasoning approach,rule and knowledge are set by experts.Aiming to the belief rule base generation without man-made error and learning algorithm with high interpretability construction,we propose a rule inference network based on rule-based inference using the evidential reasoning approach(RIMER).The the reference values always are linguistic values in RIMER.The partial derivatives of inference functions are proved as the theoretical fundamental of the proposed model.The framework and the learning algorithm of rule inference network for classification are presented.The feedforward of rule inference network using the inference process in RIMER,contributes for the interpretability.Meanwhile,parameters in belief rule base are trained by gradient descent as in neural network for belief rule base establishment.Moreover,we simplify the gradient by proposing the "fake gradient"to reduce the learning complex during the training process.We analyze the rule inference network performance in interpretability and precision from the contrast experiment results.The proposed approach reduces the confusion of interpretability benefiting from the rules and the inference engine in RIMER and constructs a rational BRB benefiting from the learning capability of neural network.In brief,this thesis proposes approximate reasoning approaches in linguistic environment based on linguistic-valued lattice implication algebra and linguistic-valued intuitionistic fuzzy lattice implication algebra respectively and an interpretability learning algorithm based on RIMER-RIN for BRB automatically generation and reducing the confusion of interpretability.The research works are significant for the research on the basic theory and interpretability of uncertain AI.