Research map / Learning theory and algorithms / Prediction and decision losses
Disentangling Linear Quadratic Control with Untrusted ML Predictions
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DISC, an online-learned, component-wise confidence policy for linear quadratic control that uses disentangled (ICA-based) ML predictions of latent perturbations, achieving near-optimal consistency and bounded robustness.
Problem
Linear quadratic control under perturbations driven by multiple latent components whose ML predictions may be unreliable; the goal is a consistency-robustness tradeoff in competitive ratio when predictions are accurate or erroneous.
Approach
Extends the single-parameter λ-confident policy to a vectorized confidence parameter per latent component, and learns it online via FTRL (linear mixing) or FTPL (general mixing) using a reduction (Lemma 1) that turns the dynamic regret into a memoryless online optimization problem, yielding a best-of-both-worlds competitive ratio.
Key results
- Theoretical result: DISC achieves CR ≤ 1 + O(sum_i ε(i)/(Ω(T/w)+ε(i))) + O(w sqrt(k/T)) + O(ρ_F^{2w} + η/T) for linear mixing, i.e., (1+o(1))-consistent and O(1)-robust with large w.
- In the drone experiment, DISC consistently yields lower cost ratio than LQR, untrusted-MPC, and self-tuning policy across scaling factors ν from 0.25 to 8.
- Online-learned confidence parameters converge to 1 for predictable latent components and decrease for unpredictable (Gaussian rain/residential) components.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| learn a confidence parameter online for each disentangled latent component instead of using a fixed confidence vector | LQR with latent perturbations and untrusted predictions · Linear mixing | competitive ratio: CR(DISC) ≤ 1 + O(∑ᵢ ε(i)/(Ω(T/w) + ε(i))) + O(w√(k/T)) + O(ρF²ʷ + η/T) | Helped |
| learn a confidence parameter online for each disentangled latent component instead of using a fixed confidence vector | LQR with latent perturbations and untrusted predictions · General mixing | competitive ratio: E [CR(DISC)] ≤ 1+O(∑ᵢ ε(i)/(Ω(T/w) + ε(i))) + O(sρF²ʷ + η/T) + O(wk²/√T) | Helped |
| use a fixed nonzero confidence vector instead of learning confidence online | λ-CON for LQR with latent perturbations · Linear mixing | competitive ratio: (1 + o(1))-consistent; at least ω(1)-robust | Hurt |
| use DISC with disentangled predictions rather than baselines that do not distinguish latent variables | Drone navigation · Drone navigation with mixed disturbances | cost ratio | Helped |
| use DISC rather than the LQR, self-tuning policy, or MPC with untrusted predictions | Voltage control · Voltage control with heterogeneous power injections | average total cost | Helped |
| reduce confidence in a latent component when its predictions become unpredictable | DISC · Voltage control with non-stationary solar generation | confidence parameter | Helped |
| increase confidence in a latent component when its predictions become more predictable | DISC · Voltage control with non-stationary wind generation | confidence parameter | Helped |
Surprising findings
- A static confidence parameter λ cannot achieve both (1+o(1))-consistency and bounded robustness for λ-CON (Theorem 4.1).
Limitations
- Disentanglement guarantees rely on independence assumptions on latent variables and ICA identifiability, which may not hold in practice.
- Theoretical analysis is for linear (and bijective Lipschitz) mixing; nonlinear dynamical systems are left to future work.
- General-mixing bound converges slower (O(k^2/sqrt(T))).
Experimental setup
Datasets: DTU solar PV data 2021 NREL US Virgin Islands wind resources synthetic drone perturbations (sinusoidal wind, Gaussian rain)
Benchmarks: drone navigation simulation (cost ratio vs offline optimal) voltage control simulation on 11-node distribution grid
Metrics: competitive ratio (cost ratio J(pi)/J*) average total cost confidence parameter convergence
Baselines: LQR MPC with untrusted ML predictions (λ=1) self-tuning policy
Builds on: λ-confident policy (Li et al. self-tuning LQC with untrusted predictions) FastICA follow-the-regularized-leader follow-the-perturbed-leader
Closest papers by meaning
- Maximizing the Value of Predictions in Control: Accuracy Is Not Enough (NeurIPS 2025 · similarity 0.85)
- The Power of Predictions in Online Control (NeurIPS 2020 · similarity 0.85)
- A Memory Efficient Unified Algorithm for Online Learning of Linear Dynamical Systems (NeurIPS 2026 · similarity 0.85)
- Safely Learning Controlled Stochastic Dynamics (NeurIPS 2025 · similarity 0.84)
- Regret Bounds for Episodic Risk-Sensitive Linear Quadratic Regulator (ICLR 2025 · similarity 0.84)
- Two‑Stage Learning of Stabilizing Neural Controllers via Zubov Sampling and Iterative Domain Expansion (NeurIPS 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.