Abstract: To improve the efficiency of the XACML(eXtensible access control markup language) policy evaluation engine under distributed environment, a novel XACML policy evaluation engine, HPEngine was proposed. The HPEngine dynamically refined policies based on statistical analysis of the policy optimization mechanism first and transformed text form of policy into numerical afterward. Moreover, the engine adopted the multi-level caching mechanism based on the statistical analysis to store frequently called request-results, attributes and policy information. Emulation results show that multi-level optimization mechanisms based on the statistical analysis applied in HPEngine significantly reduce the size of policies, decrease the communication cost between the engine and other components, lessen the amount of matching operation and improve the speed of matching. Comparative analysis demonstrates that HPEngine is obviously better in performance than other similar systems.