Traffic　congestion　is　a　common　phenomenon　in　cities,　and　improving　traffic　efficiency　has　become　an　urgent　problem　to　be　solved.　Many　researchers　have　proposed　feasible　solutions　from　different　perspectives,　such　as　traffic　flow,　signal　lights,　etc.　This　paper　proposes　a　method　of　dissipating　stop-and-go　wave　based　on　reinforcement　learning　(RL).　Specifically,　we　take　the　mixed　autonomous　vehicle　flow　as　the　research　object,　using　RL　to　train　the　driving　strategy　of　connected　autonomous　vehicle　(CAV),　and　dissipating　the　stop-and-go　waves　in　vehicle　flow　by　adjusting　the　CAV＇s　driving　behavior.　We　propose　the　concept　of　＇Equivalent　Density　Difference＇as　a　model　to　describe　the　difference　of　traffic　flow　dynamics　before　and　after　a　specific　vehicle　within　a　certain　range,　and　use　this　index　to　design　RL　model.　The　proposed　method　combines　the　advantages　of　data-driven　and　model-driven,　improving　the　training　efficiency　of　RL.　Experimental　results　show　that　this　method　can　increase　the　system-level　speed　and　improve　the　stability　of　the　mixed　autonomous　vehicle　flow.　©　2021　Technical　Committee　on　Control　Theory,　Chinese　Association　of　Automation.