*2.2. Energy Optimization Strategy*

Energy optimization in HNN is vital to ensure that every induced logic produced during retrieval phase is always achieved by global minimum energy. This creates an important question is: why HNN must achieve global minimum energy? Global minimum energy indicates a good agreement between the learned logic during pre-processing stage with the induced logic during retrieval phase. Induced logic that achieved global minimum energy can be interpreted. In contrast, induced logic that can achieve local minimum energy might achieve good accuracy, but this is difficult to interpret. In [22], the proposed logic mining is mainly the focus on the energy stability. The main issue when the induced logic is solely focusing on global minimum energy is limit on the possible search space of the HNN. The proposed HNN tends to overfit and produce more redundant induced logic. This will worsen when the proposed HNN selects the wrong attribute to learn. Non-optimal induced logic obtained a lack of interpretability and generalization during the retrieval phase. We tend to achieve similar induced logic which will lead to lower accuracy. Another factor that might affect overfitting of the induced logic structure is the monotonous behaviour of HNN that always converges to the nearest minimum energy. Hence, the feature of energy optimization with the optimal attributes selection will lead to a result that is optimal and easy to interpret.
