Source-linked AI summary

Learning Linear-Quadratic Regulators Efficiently with only $\sqrt{T}$ Regret

Alon Cohen, Tomer Koren, Yishay Mansour

arXiv:1902.06223v2cs.LGstat.ML

TL;DR

The paper addresses adaptive LQ control with unknown dynamics, where learning must occur while avoiding costly instability. It introduces an optimistic, confidence-aware semidefinite-programming algorithm and obtains approximately sqrt(T) regret efficiently, resolving prior open questions.

  • Problem

    Adaptive LQ control requires learning unknown dynamics while controlling states whose growth can make costs arbitrarily large.

  • Method

    The algorithm solves confidence-aware relaxed semidefinite programs to compute optimistic policies that balance exploitation with exploration.

  • Results

    The paper gives the first computationally efficient algorithm with a high-probability approximately sqrt(T) regret bound for learning LQ systems.

  • Takeaways & Limitations

    The result resolves open questions on achieving approximately sqrt(T) regret computationally efficiently in LQ control with unknown dynamics.

  • Takeaways & Limitations

    The analysis requires sequential strong stability, because switching among individually stable policies can otherwise cause states to grow polynomially in T.

Abstract

from arXiv · show

We present the first computationally-efficient algorithm with $\widetilde O(\sqrt{T})$ regret for learning in Linear Quadratic Control systems with unknown dynamics. By that, we resolve an open question of Abbasi-Yadkori and Szepesvári (2011) and Dean, Mania, Matni, Recht, and Tu (2018).

1 Introduction

Adaptive LQ control requires learning unknown dynamics while preventing instability and controlling regret. This paper gives the first computationally efficient algorithm with approximately sqrt(T) regret, resolving prior open problems.

  • Adaptive control must learn unknown system parameters while regulating states whose uncontrolled growth can make costs arbitrarily large.
  • Regret compares the learner’s total cost with T times the steady-state cost of the optimal policy computed using known dynamics.
  • Earlier methods achieved polynomial-dimensional regret bounds but required computationally inefficient non-convex optimization at every step.
  • Dean et al. (2018) obtained O(T^2/3) regret efficiently, while an efficient algorithm with approximately sqrt(T) regret remained open.
  • The paper presents the first computationally efficient algorithm with approximately sqrt(T) regret for learning LQ systems.
  • The approach uses optimism in the face of uncertainty, balancing exploitation of current estimates against exploration for better future estimates.

2 Preliminaries

LQ control models linear noisy dynamics with quadratic costs, while adaptive control learns unknown dynamics online. The preliminaries formulate optimal control through stability conditions and a semidefinite program.

  • LQ control: In LQ control, dynamics are linear in state and action, Gaussian noise perturbs the system, and costs are quadratic.
  • Problem setting: The learner observes each state, chooses an action, incurs a cost, and receives a successor state over T rounds.
  • Optimal policy: The optimal controllable-system policy is linear, π⋆(x) = K⋆x, and minimizes the infinite-horizon steady-state cost J⋆.
  • Stability: A policy K is stable when ρ(A⋆ + B⋆K) < 1, but stability of individual policies does not ensure safe switching between policies.
  • Problem setting: The dynamics matrices A⋆ and B⋆ are unknown, whereas the positive-definite cost matrices Q and R are fixed and known.
  • SDP formulation: The LQR planning problem can be written as a convex semidefinite program whose optimal value equals J⋆ and whose solution yields an optimal policy.
  • Stability: Sequential strong stability controls expected state magnitudes across policy sequences, preventing switching-induced state blow-up.

3 Efficient Algorithm for Learning in LQRs

OSLO learns system parameters online and computes policies through confidence-aware relaxed SDPs. The resulting policies are used for both control and data collection.

  • Algorithm structure: OSLO maintains improving dynamics estimates and a positive-definite confidence matrix, updating them in epochs when confidence-ellipsoid volume halves.
  • Parameter estimation: The algorithm uses least squares to estimate the unknown dynamics from observations and constructs confidence bounds from their covariance matrix.
  • Policy computation: It solves a relaxed semidefinite program using current estimates and confidence information to compute a stable policy.
  • Policy computation: When estimates are sufficiently accurate, the relaxed SDP underestimates the optimal steady-state cost, implementing optimism under uncertainty.
  • Control and updating: After extracting Kt, the algorithm applies ut = Ktxt, then updates its confidence matrix with the new state-action observation.
  • Control and updating: Each policy serves as both a current best guess for the optimal policy and a mechanism for collecting samples that improve later estimates.

4 Overview of Analysis

The analysis establishes a high-probability regret bound under suitable initialization, while showing that each round runs in polynomial time. Its core argument combines parameter estimation, relaxed-SDP policy computation, sequential stability, and concentration bounds.

  • Main result: Theorem 4 gives a high-probability regret bound for Algorithm 1 when its initial estimation error satisfies the stated condition.The bound applies with probability at least 1 − δ.
  • Main result: The procedure has polynomial runtime per round in the relevant problem parameters and T, log(1/δ).
  • Parameter estimation: Algorithm 1 repeatedly computes least-squares estimates of the unknown dynamics and uses confidence bounds to control estimation error.The analysis relies on a high-probability self-normalized martingale bound and on bounded visited states.
  • Policy computation via a relaxed SDP: The relaxed SDP produces policies whose extracted gains are well defined and whose objective value is controlled by the optimal steady-state cost.The SDP also supplies a positive semidefinite dual matrix used in the analysis.
  • Sequential strong stability: Repeated relaxed-SDP solutions yield a sequentially strongly stable policy sequence, ensuring the good event occurs with probability at least 1 − δ/2.The stability argument controls state growth across policy updates rather than only for each policy separately.
  • Regret analysis: The regret proof combines SDP-based instantaneous-regret bounds, martingale concentration, telescoping across unchanged policies, and at most O(n log T) policy switches.The main theorem follows by combining these bounds with a union bound and algebraic simplification.

5 The relaxed SDP program

The paper repeatedly solves a relaxed semidefinite program that approximates the exact LQ control program and extracts deterministic linear policies. As the relaxations tighten, stability properties of the resulting policy sequence support the analysis.

  • Relaxed SDP formulation: The relaxed program imposes a matrix inequality involving the estimated dynamics, process noise, and confidence matrix V.The dynamics error is constrained through Tr(∆V∆^T) ≤ 1.
  • Relaxed SDP formulation: The relaxed SDP is a relaxation of the exact SDP when the dynamics estimates are sufficiently accurate.Its optimal value is at most the optimal infinite-horizon cost, with bounded primal and dual solution norms.
  • Policy extraction: The SDP yields a deterministic linear policy K = Σ_uxΣ_xx^-1, provided Σ_xx is invertible.The analysis establishes positive definiteness of Σ_xx using primal-dual properties and rank arguments.
  • Sequential stability: The sequence of policies generated by the relaxed SDPs is strongly stable under the stated conditions on V_t and µ.Theorem 17 requires V_t ⪰ 16κ^10µI, while Lemma 18 supplies uniform bounds on the associated factors.
  • Sequential stability: The optimal SDP solutions remain close across rounds, helping ensure that successive extracted policies do not change drastically.Lemma 19 gives the comparison P_t ⪯ P⋆ ⪯ P_{t+1} + (α_0γ/2)I.

6 Warm-up Using a Stable Policy

The warm-up phase uses a known strongly stable policy with Gaussian exploration to collect samples and initialize the confidence matrix. A high-probability theorem provides the matrix bounds needed by the main algorithm.

  • Warm-up procedure: The warm-up procedure assumes a known policy K_0 that is (κ_0, γ_0)-strongly stable.Starting from x_1 = 0, it runs for T_0 rounds under exploratory Gaussian actions.
  • Warm-up procedure: Gaussian exploration actions are sampled independently around K_0x_t with covariance 2σ^2κ_0^2I.The collected vectors z_t = (x_t,u_t) determine the empirical covariance used to initialize V_0.
  • Warm-up guarantee: The warm-up analysis adapts identification techniques for linear dynamical systems and concentration inequalities for Gaussian and martingale quantities.The proof invokes Hanson-Wright, Azuma, and self-normalized concentration tools.

A.3 Proof of Lemma 3

The proof bounds the state and regret processes along the sequentially strongly stable policy sequence. It combines confidence-matrix control, determinant updates, and concentration arguments to obtain the theorem’s high-probability conclusion.

  • State control: Sequential strong stability ensures that expected state-vector magnitudes remain controlled.The dynamics are represented through uniformly bounded similarity transforms and contracting factors.
  • State control: The proof introduces events E_t under which parameter estimates are accurate and state-feature norms remain bounded.These events are used to control the confidence matrices and associated regret quantities.
  • Confidence and updates: Matrix norm bounds and the relaxed-SDP properties control the per-round regret contributions on the good events.The argument uses bounds on ∥z_t∥ and ∥P_t∥_* together with Lemma 27.
  • Confidence and updates: The algorithm updates its policy when the determinant of the confidence matrix grows sufficiently, yielding a bounded number of policy changes.The proof uses det(V_t) ≤ 2 det(V_τ(t)) between update times.
  • Conclusion: A union bound and algebraic simplification complete the high-probability regret argument.The final step plugs in the bounds from Lemmas 10–13.

B.2 Proof of Corollary 5

The proof of Corollary 5 transfers the warm-up covariance guarantee into the initial conditions required by the main theorem. It then combines the theorem and auxiliary lemma through a union bound to bound total regret.

  • Initialization: The lower bound V ⪰ V_0 ⪰ (T_0σ^2/80)I makes the initial conditions of Theorem 4 hold when Theorem 20 succeeds.Thus, the warm-up duration supplies the required initialization scale.
  • High-probability guarantee: With probability at least 1 − δ, Theorems 4 and 20 and Lemma 6 hold simultaneously after allocating failure probability across them.The proof applies a union bound with probability at least 1 − δ/3 for each component.
  • High-probability guarantee: The total regret combines the warm-up regret with the remaining-round regret controlled by Theorem 4.The proof explicitly separates the initial exploration phase from the subsequent main-algorithm phase.
  • Dynamics estimation: The dynamics estimator is analyzed by writing x_{s+1} = Θ⋆z_s + w_s and applying self-normalized concentration coordinatewise.A union bound then makes the coordinatewise bounds hold simultaneously.
  • Dynamics estimation: The resulting estimation bounds ensure Tr(∆_tV_t∆_t) ≤ 1 under Algorithm 1’s conditions.This verifies the confidence condition needed by the relaxed-SDP analysis.

B.4 Proof of Lemma 7

The proof establishes the lemma by applying Lemmas 15 and 16 to bound the SDP quantities and connect them to the true parameters.

  • Lemma 15 gives item (i) and bounds the SDP dual solution as ∥P_t∥_* ≤ ν/σ2.
  • V_t is lower bounded by (νμ)/(α0σ2) times the identity, enabling Lemma 16 to establish item (ii).
  • The dual matrix P_t is positive semidefinite by the SDP dual formulation, while Lemma 16 supplies an additional bound for item (iii).
  • The proof links the resulting equation to the true parameters (A*, B*) by combining it with Lemma 24.

B.5 Proof of Lemma 8

The proof controls the algorithm’s policies, states, and SDP quantities on a high-probability event, using concentration inequalities, stability, and policy-update counting.

  • A union bound combines Lemma 6 and Hanson-Wright concentration to obtain simultaneous validity with probability 1 − δ.
  • The event ∥V_t∥ ≤ 4T ensures sequential strong stability and yields a bound on the state norms ∥z_t∥2.
  • Theorem 17 applies to the policy sequence once μ ≥ 1 + 2ϑ∥V_t∥1/2 and V_t ⪰ 16κ10μI hold for every round.
  • Algorithm 1 changes policy at most 2n log T times under the event E_N.
  • A Gaussian martingale concentration argument bounds sums involving the SDP matrices by truncating quadratic forms and applying Azuma and Hanson-Wright inequalities.
  • Applying the resulting concentration lemma with M_t = P_t I{E_t} uses the bound ∥P_t∥_* ≤ ν/σ2 on E_t.

C Proofs of Section 5

These proofs compare the SDP solution with the optimal Riccati solution and use matrix inequalities to establish ordering and norm properties of the resulting policies.

  • Lemmas 16 and 24 provide the principal matrix inequalities used to compare the SDP quantities with the optimal solution.
  • The assumption V ⪰ 2κ2μI allows Lemma 14 to control the policy-related quantities.
  • The Riccati equation characterizes P*, whose associated optimal policy K* is strongly stable.
  • Subtracting the derived matrix inequalities establishes the required ordering between P and P* and supports the final norm bound.

D Proofs of Section 6

The proofs establish high-probability state and covariance bounds, lower-bound the smallest eigenvalue of the covariance matrix, and use these properties to control estimation and regret arguments.

  • Lemma 32 provides a high-probability uniform bound on the states over t = 1, . . . , T0 + 1.
  • Under the strongly stable policy K0, the state process is analyzed as an LQR system with modified Gaussian noise B*η_t + w_t.
  • For any unit vector n, Lemma 33 gives n^T V n ≥ T0 · σ2/40 when T0 ≥ 200 log(1/δ).
  • The proofs use Gaussian tails, martingale differences, and Azuma’s inequality to control indicator events and their cumulative behavior.
  • The main theorem proof begins by introducing Gaussian exploration noise η_t with variance 2σ2κ2 and then applies the preceding bounds.
  • A 1/4-net argument yields ∥V^-1∥ ≤ 80/(T0σ2), provided T0 ≥ 400(n + log(1/δ)).
Loading 1902.06223v2…