基于深度强化学习PPO的匝道混合交通流合流控制方法OA

[Background]The gradual integration of autonomous vehicles(AVs)into traffic systems has increased the prevalence of mixed traffic flows,which comprise human-driven vehicles(HDVs)and AVs in on-ramp merging areas.However,single-vehicle intelligence approaches exhibit limita-tions in terms of real-time responsiveness and system-wide coordination.Moreover,the overall char-acteristics and comprehensive effects of mixed traffic flows require further in-depth investigation.[Objective]This study aims to address control challenges posed by mixed traffic flows in on-ramp merging areas to enhance the overall traffic efficiency and safety.[Method]A control method for AVs is proposed by leveraging the proximal policy optimization(PPO)deep reinforcement-learning(DRL)algorithm to execute car-following and lane-changing behaviors.To mitigate frequent lane changes and prevent overly conservative driving,this strategy incorporates a lane-changing penalty and a fixed-step negative reward mechanism.Simulations are conducted to evaluate the effect of DRL-based AVs on mixed traffic flows.[Result]As the AV penetration rate increases,the proposed strategy significantly enhances both traffic efficiency and safety.Compared with a DRL baseline(DRL-B)and a rule-based(RB)strategy,this method improves the overall traffic efficiency and two safety indicators,TTC and DRAC,by 2.31％,17.3％,3.1％,and 4.57％,10.7％,0.34％,respectively.The most significant improvement in operational efficiency is observed under moderate arrival rates when compared with the RB strategy,and under high arrival rates when compared with the DRL-B strategy.The most substantial safety improvements are observed at moderate AV penetration rates.[Application]This study validates the effectiveness of the DRL approach in enhancing the efficiency and safety of mixed traffic flows in on-ramp merging areas,as well as offer insights for the deploy-ment and application of AVs in such scenarios.