We propose a method to obtain correct inference rules from experience for an intel- ligent agent running in an environment. For each inference rule, the proposed method substitutes "efficiency to reach the correct answer" for "correctness." This proposed method learns the infer- ence rules using a hierarchical reinforcement learning method called RGoal. The whole architecture is biologically plausible. We believe the proposed method will be a basic principle of autonomous knowledge acquisition for artificial general intelligence.