Accurate　decision-making　within　highly　interactive　driving　environments　is　vital　for　the　safety　of　self-driving　vehicles.　Despite　the　significant　progress　achieved　by　the　existing　models　for　autonomous　vehicle　decision-making　tasks,　there　remains　untapped　potential　for　further　exploration　in　this　field.　Previous　models　have　focused　primarily　on　specific　scenarios　or　single　tasks,　with　inefficient　sample　utilization　and　weak　robustness　problems,　making　them　challenging　to　apply　in　practice.　Motivated　by　this,　a　robust　decision-making　method　named　DRL-EPKG　is　proposed,　which　enables　the　simultaneous　determination　of　vertical　and　horizontal　behaviors　of　driverless　vehicles　without　being　limited　to　specific　driving　scenarios.　Specifically,　the　DRL-EPKG　integrates　human　driving　knowledge　into　a　framework　of　soft　actor-critic　(SAC),　where　we　derive　expert　policy　by　a　generative　model:　variational　autoencoders　(VAE),　train　agent　policy　by　employing　the　SAC　algorithm　and　further　guide　the　behaviors　of　the　agent　by　regulating　the　Wasserstein　distance　between　the　two　policies.　Moreover,　a　multidimensional　reward　function　is　designed　to　comprehensively　consider　safety,　driving　velocity,　energy　efficiency,　and　passenger　comfort.　Finally,　several　baseline　models　are　employed　for　comparative　evaluation　in　three　highly　dynamic　driving　scenarios.　The　findings　demonstrate　that　the　proposed　model　outperforms　the　baselines　regarding　the　success　rate,　highlighting　the　practical　applicability　and　robustness　of　DRL-EPKG　in　addressing　complex,　real-world　problems　in　autonomous　driving.