A Decision-Making And Verification Approach For Responsible Agents In Autonomous Driving Systems

With the development of trusted AI,how to make the agent’s decision-making process follow the human-like thinking mode and ensure its responsible behavior is one of the challenging problems to delve into autonomous driving technology.But the definition of responsible behaviors in existing work is still not detailed enough.In addition,because the decision in the autonomous-driving scenario is extremely sensitive to uncertain factors,the probability-reasoning decision method based on Bayesian network can solve this problem very well.However,the existing Bayesian network decision-making methods are still powerless to generate responsible behaviors of agents; Moreover,there is still a lack of an appropriate logical language and method to specify and verify such responsible behaviors in the existing work.In view of the above problems,we first give the specific meaning of the responsible behaviors of agents; And the BDI model is used to guide the decision-making process of Bayesian network to improve its decision confidence; Then,the specification language BDI-XSTIT is designed to specify the responsible behaviors of agents,and the modelchecking algorithm based on BDI-XSTIT is given; The Py Nu SMV tool is used to realize automatic verification based on BDI-XSTIT; Finally,the rationality and effectiveness of the approach are demonstrated by the experiments in the lane-change and overtaking scenario of autonomous driving.Specifically,the main work of this paper is as follows:· In order to realize the responsible decision-making of autonomous driving,a BDI model-based Bayesian network decision-making approach is proposed.Firstly,the meaning of responsible behavior is defined,that is,the external behaviors are generated under the guidance of internal behaviors; Then,based on domain knowledge,the BDI model in the autonomous-driving scenario is analyzed,and the intention inference algorithm is designed as the internal behavior reasoning module; Combined with LSTM model to predict the intention of other vehicles,the BDI-hierarchical Bayesian network facing the autonomous-driving scenario is constructed through structural learning and parameter learning,so that its decisionmaking process is guided by the internal behaviors of the agent on the basis of traditional probabilistic reasoning,improving the confidence of decisions,and achieving responsible decision making.· In order to solve the problem of verification based on responsible behaviors of agents,the logic language BDI-XSTIT for specifying responsible behaviors is proposed,and the model-checking algorithm based on BDI-XSTIT is designed.Firstly,the formal syntax and semantics of BDI-XSTIT logic are defined,which can describe the mental states(belief,desire,intention),agency and knowledge of the agent; Then we discuss the properties of BDI-XSTIT itself,such as the relations between the mental states,agency and knowledge and so on; In addition,the model-checking problem based on BDI-XSTIT is explored,and the definition of finite-dynamic-state structure is proposed.Each possible world and the cognition of agents related to BDI are modeled as Kripke structures,and the dynamic changes between possible worlds are modeled by accessibility relations.The PSPACE decision algorithm is given,and the Boolean predicate is used for coding implementation; Finally,the model-checking problem based on BDI-XSTIT is realized with the Py Nu SMV tool.· We take the lane-change and overtaking scenario as an example,and design an experiment with the approach of responsible decision-making and verification proposed in this paper.We introduce the application of Bayesian network based on BDI model in lane-change and overtaking scenario and the improvement of decision confidence; We also introduce how to use BDI-XSTIT logic to specify agents’ responsible behaviors,describe the implementation process of its model-checking in detail,and show the rationality and effectiveness of the approach proposed in this paper.