Self-learning adaptive neuro-fuzzy approximation of robust control behavior in electric power steering systems

Area of Science:

• Artificial Intelligence
• Machine Learning
• Control Systems

Background:

• Artificial Intelligence (AI) data training algorithms frequently face challenges such as overfitting, underfitting, and bias.
• Existing methods may struggle with accuracy and generalization, especially in dynamic or novel scenarios.

Purpose of the Study:

• To design and evaluate a hybrid self-learning algorithm that addresses common data training issues in AI.
• To improve the accuracy and generalization capabilities of AI training algorithms.

Main Methods:

• Developed a hybrid algorithm by integrating fuzzy logic and neural network structures into an Adaptive Network-Based Fuzzy Inference System (ANFIS).
• The ANFIS model was designed with three inputs and one output.
• Trained the model using data from a high-performance robust controller for Electric Power Steering (EPS) systems.
• Compared the proposed ANFIS against the Backpropagation Neural Network (BPNN) as a benchmark.

Main Results:

• The proposed ANFIS demonstrated high training accuracy with errors below 1.7% in well-trained cases.
• Maintained strong interpolation capabilities with errors under 6.1%.
• Achieved prediction errors of less than 9.3% for scenarios outside the training domain.
• Significantly resolved overfitting issues compared to the benchmark BPNN.

Conclusions:

• The hybrid ANFIS algorithm offers superior data training accuracy and generalization performance.
• This approach effectively mitigates overfitting, outperforming conventional methods like BPNN.
• The ANFIS model shows promise for robust control applications in systems like Electric Power Steering.