TL;DR
This paper introduces a "task-aware" data selection method for Nonlinear Data-Enabled Predictive Control (DeePC). By leveraging a machine learning framework called "datamodels," the authors train a neural network to predict which specific pieces of historical data will best help a controller achieve its current goal, leading to significantly better tracking performance with much less data.
What Problem Does This Paper Address?
Data-Enabled Predictive Control (DeePC) is a "model-free" approach to automation. Instead of building a complex mathematical model of a system (like a robot arm or a chemical plant), DeePC uses a big matrix of historical input-output data (a Hankel matrix) to predict future behavior.
While this works beautifully for linear systems, real-world systems are often nonlinear. To handle this, researchers usually try to find "local" data—historical snippets that look like the robot's current state—to create a temporary linear approximation. This is called geometric proximity selection.
However, the authors point out a major flaw: geometric proximity ignores the goal. If a robot is at a standstill, the data it needs to perform a sharp, aggressive jerk is very different from the data it needs to perform a slow, smooth glide. Current methods can't tell the difference because they only look at where the robot is, not where it's going.
Methods
The authors propose a Context-Dependent Linear Datamodel. Their approach is inspired by recent breakthroughs in machine learning interpretability that show the output of complex models can often be approximated by a simple linear function of their training data.
- The Datamodel: They treat the "closed-loop cost" (a measure of how poorly the controller tracks a target) as a linear sum of "influence scores" assigned to each column of data in the Hankel matrix.
- Context-Awareness: Unlike previous work, these influence scores aren't static. They are generated by a Neural Network (a Multi-Layer Perceptron) that takes the "context"—the current initial trajectory and the future reference trajectory—as input.
- Offline Training: To teach the network, they run thousands of closed-loop simulations using random subsets of data. The network learns to correlate specific data columns with success or failure in specific scenarios.
- Online Selection (Top-K): During real-time operation, the network does a quick "forward pass" to score all available data columns. The controller then picks the Top-K columns with the best (most negative) influence scores and solves a much smaller, faster optimization problem.
Datasets
The researchers evaluated their method using the Reacher environment, a standard benchmark involving a planar two-degree-of-freedom robotic manipulator.
- Data Volume: They collected 200 simulations of 200 steps each, resulting in a Hankel matrix with 36,800 columns.
- Training: They generated a training dataset of 50,000 samples via Bernoulli column sampling to train the influence network.
- Baselines: They compared their results against $L_1$ selection (geometric proximity), Isomap (manifold embedding), and random sampling.
Key Insights and Ideas
- The "Aha!" Moment: Even though the underlying physics and the control optimization are highly nonlinear and complex, the utility of a specific data point can be modeled linearly. This makes the selection process computationally efficient enough for real-time use.
- Task-Awareness is King: The results show that knowing the destination is just as important as knowing the starting point. By including the reference trajectory in the "context," the controller can pick data that captures the specific dynamics required for the upcoming maneuver.
- Efficiency through Sparsity: The method performs exceptionally well even when using a tiny fraction of the available data (less than 0.5%). This suggests that most historical data is "noise" for any specific task, and finding the "signal" is the key to performance.
Key Results
- Superior Tracking: The datamodel-based approach consistently achieved the lowest tracking cost across all tested metrics (Quadratic Cost, IAE, and ISE).
- 2x Performance Boost: At small subset sizes (100–200 columns), the proposed method achieved approximately 2x lower cost than geometric baselines like $L_1$ selection.
- Computational Savings: By selecting a small "Top-K" subset, the complexity of the quadratic programming (QP) solver is reduced from $O(M^3)$ to $O(K^3)$, where $K \ll M$. This makes DeePC much more viable for systems with massive historical datasets.
Limitations and Future Work
- Offline Dependency: The current framework requires a significant amount of offline simulation to train the influence network. If the system's dynamics change (e.g., a motor wears out), the network might need retraining.
- Fixed Alpha: Currently, separate networks are trained for different target subset sizes. The authors suggest a "Unified Model" that can generalize across any subset size $K$ as a future improvement.
- Online Adaptation: A promising next step is "Online Model Adaptation," where the influence scores are updated incrementally as new data arrives, allowing the controller to learn and adapt to drifting dynamics on the fly.
Summary updated