TL;DR
This paper introduces a novel Iterative Learning Control (ILC) framework that leverages machine learning to tackle repetitive tasks in systems with non-repetitive, time-varying dynamics. By using ML to continuously update a nominal model of the system between trials, the approach significantly improves tracking accuracy and expands the applicability of ILC to a broader class of challenging real-world systems.
What Problem Does This Paper Address?
Traditional Iterative Learning Control (ILC) excels at improving the performance of systems performing the exact same task repeatedly. However, a key assumption in classical ILC is that the system's dynamics remain constant across all trials. This assumption breaks down in many real-world scenarios where system parameters drift or change over time, even within a repetitive task. These "non-repetitive time-varying systems" (NTVSs) pose a significant challenge because the accumulated model uncertainties can lead to performance degradation and even instability in standard ILC algorithms. Existing methods that handle time-varying systems often still assume the dynamics are repetitive across trials, or they require bounded system parameters, which is not always the case for NTVSs. This paper aims to bridge this gap by developing an ILC approach that can adapt to and effectively control NTVSs.
Methods
The core innovation lies in the integration of a machine learning (ML) based nominal model update mechanism within an ILC framework.
- ML-Based Nominal Model Update: At the end of each ILC trial, during a brief initialization period, a linear regression technique (a form of ML) is employed. This ML module uses the input, output, and state data from the current trial to learn and update a nominal model of the system's dynamics. This updated nominal model is then used for the next ILC trial. The key idea is that this ML update forces the model uncertainties to remain within a tolerable bound for the ILC algorithm.
- Feedforward ILC Update Law: A modified norm-optimal ILC update law is proposed. This law incorporates the ML-estimated nominal model and also accounts for the changes in the system dynamics between adjacent trials. This is crucial for NTVSs where the system is not static. The update law aims to minimize both the tracking error and the change in the control input between trials, promoting robustness.
- Parameter Tuning: The paper provides guidance on how to tune the parameters of both the ML and ILC components to achieve the desired overall performance. This includes setting weighting matrices for the ILC and learning rates for the ML gradient descent.
Datasets
The paper does not explicitly mention the use of pre-existing datasets for training or evaluation. Instead, it relies on simulated data generated from a specific NTVS model (a gantry robot) for its numerical case study. The "dataset" for the ML training is generated on-the-fly by recording input, output, and state values during each ILC trial.
Key Insights and Ideas
- Breaking the Invariance Postulate: The most significant contribution is the successful removal of the ILC design postulate that system dynamics must be invariant across trials. This opens up ILC to a much wider range of practical applications.
- ML as a Dynamic Model Estimator for ILC: Using ML, specifically linear regression, to continuously estimate and update the system's nominal model between ILC trials is a novel and effective way to handle non-repetitive time-varying dynamics. This allows the ILC to adapt to changes without requiring a perfect, pre-defined model.
- Robustness through Bounded Uncertainties: The ML update mechanism is designed to keep model uncertainties within a bound that the ILC algorithm can handle, ensuring robust convergence even with changing system dynamics.
- Trade-off between Accuracy and Input Change: The proposed ILC update law balances minimizing tracking error with minimizing input changes between trials, a common strategy for ensuring robustness and preventing excessive control effort.
Key Results
The proposed algorithm was validated through a numerical simulation case study using a gantry robot model with embedded non-repetitive time-varying parameters.
- Superior Tracking Accuracy: The algorithm achieved a mean absolute tracking error significantly lower (on the order of 10⁻¹ to 10⁻² mm) than classical ILC and a standard feedback control method (which achieved errors between 10⁻¹ and 10⁰ mm).
- Convergence to Upper Bound: The tracking error converges to a small, bounded level, as predicted by the theoretical analysis (Theorem 1). While it doesn't reach zero due to the inherent time-varying nature, it achieves a level of accuracy close to the hardware limitations (optical encoder resolution).
- Effective Model Adaptation: The ML component successfully estimated the time-varying parameters of the system, keeping the model uncertainties within a tolerated bound and enabling the ILC to perform effectively.
- Reasonable Input Signals: Even with a high weighting on error reduction (large Q), the input signals at the final trial remained reasonable.
Limitations and Future Work
- Theoretical Convergence to a Bound: The paper acknowledges that for NTVSs, the tracking error is guaranteed to converge to a small, bounded interval, not necessarily zero, due to the time-varying nature of the system.
- Assumption on Dynamics Change Rate: Theorem 1 relies on the assumption that the change in system dynamics between adjacent trials is not too large. If dynamics change drastically, performance might be affected.
- Observability: The current ML approach assumes system states are observable. The authors suggest that for unobservable systems, a neural network model with two layers could be used, which is noted as a direction for future work.
- Practical Test Platform: The authors propose establishing a practical test platform to validate the algorithm's performance on real hardware.
- Point-to-Point Tracking and Multi-Agent Systems: Future work could extend the approach to point-to-point tracking tasks and leader-follower formation control problems for multi-agent NTVSs.
Summary updated