REVIEW 4 major objections 5 minor 2 cited by
Coreset-Based Task Selection for Sample-Efficient Meta-Reinforcement Learning
T0 review · 4 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash
Pith's one-line read A weighted coreset of the most gradient-diverse tasks reduces MAML-RL sample complexity to an epsilon-stationary point by O(1/epsilon), and MAML-LQR by O(log(1/epsilon)).
desk verdict The main theorem's bias bound has an invalid max/Σ step; the paper's central sample-complexity claim is not proven, but the algorithm and experiments are worth a look. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The load-bearing object is the facility-location submodular function $F(\mathcal{S}) = C - \sum_{j\in\mathcal{M}}\min_{i\in\mathcal{S}}\|g_j(\theta_0)-g_i(\theta_0)\|$, computed from two-point zeroth-order estimates $g_i(\theta_0)$ of the task-specific meta-gradients. Maximizing $F$ with $|\mathcal{S}|\le L$ chooses the tasks whose gradients are closest to all others, and the greedy algorithm inherits the $(1-e^{-1})$ approximation bound of submodular maximization. Each selected task receives weight $\gamma_i = \sum_j \mathbf{1}\{j = \arg\min_{T_i\in\mathcal{S}}\|g_j-g_i\|\}$, so the weighted sum over $\mathcal{S}$ approximates the full-pool gradient. The argument then controls $\|\nabla J(\theta)-\nabla_\mathcal{S} J(\theta)\|$ as the sum of a zeroth-order estimation error, bounded by matrix Bernstein concentration, and a selection bias, bounded by the greedy submodular guarantee; under $C=O(\epsilon)$ the whole error is $O(\epsilon)$.
What would settle it
On a concrete task pool, form a coreset with $L=O(1)$ and measure the residual bias $\frac{1}{M}\sum_{j\in\mathcal{M}}\min_{i\in\mathcal{S}^\star}\max_{\theta\in\Theta}\|\nabla J_i(\theta)-\nabla J_j(\theta)\|$ using the paper's zeroth-order estimates over a grid of $\theta$. A pool where this bias is not $O(\epsilon)$ would fall outside Corollary 1's scope; a pool where the bias is $O(\epsilon)$ but the empirical sample ratio $S^M_c/S^S_c$ does not grow like $1/\epsilon$ would indicate the claimed mechanism is not what drives the speedup.
Extended reading notes
Core claim
The paper establishes that a weighted coreset of tasks can replace the full task pool in the MAML-RL update without changing the final convergence guarantee, up to a bias term $\frac{1}{M}\sum_{j\in\mathcal{M}}\min_{i\in\mathcal{S}^*}\xi_{i,j}$ that is assumed to be $O(\epsilon)$. Theorem 1 shows that Algorithm 1 reaches an $\epsilon$-close stationary solution after $N = O(1/\epsilon)$ iterations for non-concave rewards, and Corollary 1 gives the sample-complexity reduction $S^M_c = O(1/\epsilon)S^S_c$ when $M = O(1/\epsilon)$ tasks are available and the coreset size is $L = O(1)$. In the LQR setting, Theorem 2 and Corollary 2 upgrade the iteration count to $N = O(\log(1/\epsilon))$ using gradient dominance, giving a sample reduction proportional to $O(\log(1/\epsilon))$. The proof separates the gradient error into a zeroth-order estimation error, controlled by matrix concentration, and a task-selection bias, controlled by submodular maximization, so the saving persists regardless of the derivative-free estimator used.
Load-bearing premise
The load-bearing premise is that the task pool is redundant enough for a constant-sized subset to approximate every task's gradient to within error $O(\epsilon)$; if the tasks are genuinely heterogeneous, the claimed $O(1/\epsilon)$ sample saving does not follow.
Editorial extensions
If this is right
- On a large pool of similar tasks, training MAML on a constant-size weighted coreset reaches an $\epsilon$-stationary point after $N = O(1/\epsilon)$ iterations using $S^S_c = O(L/M \cdot d^2/\epsilon^3) + O(d^2/\epsilon^2)$ samples, versus $S^M_c = O(M \cdot d^2/\epsilon^3)$ samples without selection; the ratio is $O(1/\epsilon)$ when $M = O(1/\epsilon)$.
- In the LQR case, gradient dominance upgrades the iteration count to $N = O(\log(1/\epsilon))$; the sample saving over the full pool is then a factor $O(\log(1/\epsilon))$ rather than $O(1/\epsilon)$.
- The selected coreset remains stabilizing in MAML-LQR throughout training; task selection does not sacrifice the guarantee that every iterate is a stabilizing controller.
- The coreset benefit is conditional: in a heterogeneous task regime, a small $L$ can prevent convergence to a stationary solution because the selection bias term $\frac{1}{M}\sum_{j\in\mathcal{M}}\min_{i\in\mathcal{S}^*}\xi_{i,j}$ is not small.
Reading between the lines
- Beyond the paper: because the bias term is the only place task similarity enters, a practical adaptive rule suggests itself: increase $L$ until the measured facility-location residual falls below the target $\epsilon$; the theorem then predicts the sample saving on any given pool without assuming similarity in advance.
- Beyond the paper: the same decomposition of zeroth-order error plus selection bias transfers to any meta-learning update that is a sum of task gradients, so coreset selection by gradient diversity should give analogous sample-complexity gains for Reptile-style or federated meta-learning.
- Beyond the paper: a direct test of the causal mechanism would be to measure the actual bias $\frac{1}{M}\sum_{j\in\mathcal{M}}\min_{i\in\mathcal{S}^*}\xi_{i,j}$ on the MuJoCo pools used in the paper; the theory predicts it is $O(\epsilon)$ exactly where the reported speedup appears, which would distinguish a selection-driven gain from a batch-size artifact.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper proposes a coreset-based task selection method for model-agnostic meta-reinforcement learning (MAML-RL). The method selects a weighted subset of tasks by greedy submodular maximization of a facility-location objective defined on derivative-free gradient estimates at the initial parameter θ0, then runs zero-order MAML on the selected subset. The authors claim an ergodic convergence rate N = O(1/ε) for general non-concave rewards (Theorem 1), a sample-complexity reduction of O(1/ε) relative to training on the full task pool (Corollary 1), and analogous results for MAML-LQR with N = O(log(1/ε)) and a sample-complexity reduction of O(log(1/ε)) (Theorem 2 and Corollary 2). Numerical experiments on cartpole, hopper, walker2D, and LQR are presented as supporting evidence. The central theoretical quantity is an additive task-selection bias that the authors attempt to control via the greedy submodularity guarantee and a small-constant assumption (Assumption 3).
Significance. If the central proof were valid, this would be a useful contribution: it is, to my knowledge, the first derivative-free coreset task-selection scheme for MAML-RL with convergence and sample-complexity guarantees, and the LQR case study with stability analysis is a valuable extension of the MAML-LQR line of work. The paper also ships clearly structured concentration-of-measure arguments for the zero-order gradient estimators and a careful accounting of the coreset selection overhead. However, the main advertised reduction is conditional in a way that the formal statements do not fully acknowledge, and the proof of the key bias bound contains an invalid uniformization step. The numerical results are suggestive but do not by themselves establish the sample-complexity claim, since the experiments compare wall-clock or iteration progress rather than the total sample count required to reach a fixed ϵ tolerance. I therefore view the contribution as potentially salvageable but not established in its current form.
major comments (4)
- [Appendix 6.3] The proof of the task-selection bias bound is invalid at the step where the greedy guarantee is converted into a uniform-in-θ bound. The Nemhauser guarantee F(S) ≥ (1 − e^{−1})F(S⋆) holds for the facility-location objective evaluated at the single selection point θ0; it is an inequality at θ0 and does not imply the displayed inequality with max_θ inside the sum and min after 'taking the maximum of both sides', because both S and S⋆ are defined with respect to θ0, not with respect to the θ that maximizes the left-hand side. A concrete failure: take Θ=[0,1], tasks g1(θ)=100θ, g2(θ)=g3(θ)=θ, L=1, θ0=0. All gradients coincide at θ0, so C=0 and Assumption 3 holds for every ε>0. The greedy choice S={1} has true bias (1/3)Σ_j min_{i∈S}∥g_j(1)−g_i(1)∥=66, while the claimed right-hand side with S⋆={2} is (1−e^{−1})·33≈20.9. Hence the bound ∥∇MJ(θ)−∇SJ(θ)∥ ≤ ε + (1/M)Σ_{j∈M} min_{i∈S⋆} ξ_{i,j}, which is used in Theorem 1, is not established.
- [Assumption 3 and Eq. (4)] C in Eq. (4) is not a freely tunable constant. For F(S) to be a nonnegative submodular function and for the greedy guarantee to apply, C must be at least max_{S:|S|≤L} Σ_{j∈M} min_{i∈S} ∥g_j(θ0)−g_i(θ0)∥. Setting C=O(ε) is therefore equivalent to assuming that a constant-size coreset achieves facility-location error O(ε) at θ0, which is essentially the task-selection bias condition needed for Corollary 1. The subsequent requirement C ≤ M e ε / Capprox,1 in §6.3 is compatible only when the task pool is highly redundant. This should be stated explicitly as a similarity assumption on the task distribution rather than as a tuning choice; the paper's own discussion in Section 3.1 concedes that small L can prevent convergence when tasks are sufficiently different, but the formal statements do not carry this qualification.
- [Corollary 1 and Appendix 6.5] The claimed O(1/ε) sample-complexity reduction is largely a consequence of the definition S^S_c = L/M S^M_c + O(M ns) rather than of any property of the selected tasks. Since S^M_c = O(M N ns) and N = O(1/ε), any choice with L=O(1) and M=O(1/ε) yields a ratio O(1/ε) whenever the iteration count is unchanged; the only substantive requirement is that the bias term in Theorem 1 be O(ε) with L=O(1), which is exactly what the proof does not establish. In addition, the proof line in §6.5 stating 'S^M_c = O(M N ns) = O(d^2/ε^3)' drops the factor M; with M=O(1/ε), the full-pool complexity is O(d^2/ε^4). The ratio can be repaired arithmetically, but as written the proof is inconsistent.
- [Theorem 2 and Corollary 2] Theorem 2 and Corollary 2 inherit the invalid task-selection bias bound from Appendix 6.3, so the LQR guarantees are subject to the same objection. Moreover, the convergence rate in Theorem 2 is the same as the no-selection MAML-LQR rate plus a heterogeneity bias, so the sample-complexity reduction in Corollary 2 again rests on the accounting identity S^S_c = L/M S^M_c + O(M ns) rather than on an improved iteration complexity. The proof of Corollary 2 in §6.7 also writes ¯S^M_c = O((d^2/ε^2) log(1/ε)) without the factor M; with N=O(log(1/ε)) and ns=O(d^2/ε^2), the correct expression is O(M(d^2/ε^2) log(1/ε)). The final ratio may still be O(log(1/ε)) after substituting M=O(log(1/ε)), but the displayed derivation is incorrect.
minor comments (5)
- [Appendix 6.3] In the final displayed inequality of the task-selection subsection, the term ∥∇MJ(θ) − ∇MJ(θ)∥ on the left should be ∥∇MJ(θ) − ∇SJ(θ)∥.
- [Assumption 1] Assumption 1 contains the duplicated phrase 'for for any θ, θ′ ∈ Θ'.
- [Appendix 6.8] In the stability analysis, the expression 'f (ϵtext)' appears where the heterogeneity bias f(ϵhet) is meant; this should be corrected.
- [Figure 3] The left panel of Figure 3 labels the vertical axis 'T ask-Specific Optimality Gap' with a stray space; it should read 'Task-Specific Optimality Gap'.
- [Remark 1] Remark 1 is a substantial caveat: the theory assumes an oracle for expected task rewards J_j(θ), while the experiments use empirical rewards. The authors state that the results can be extended by taking nτ large enough, but this caveat should appear in the main text, not only in the appendix, because it affects the practical interpretation of the sample-complexity claims.
Circularity Check
The advertised O(1/ε) sample-complexity reduction is a definitional consequence of the assumed task-pool/coreset sizes; the coreset bias bound that would make it substantive is not derived.
-
self definitional
[Section 3.1, Corollary 1 (proof in Section 6.5)]
"Let S S_c := L/M S M_c + O(M ns) and S M_c := O(M N ns) denote the total number of samples in Algorithm 1 to find an ϵ-near stationary solution, with and without task selection, respectively. Corollary 1. ... Suppose the number of iterations scales as N = O(1/ϵ) and the number of tasks in the task pool is sufficiently large as M = O(1/ϵ). Therefore, our coreset-based task selection offers a sample complexity reduction such as S M_c = O(1/ϵ)S S_c."
Substituting the paper's own definitions gives S M_c/S S_c = M N ns / ((L/M)M N ns + M ns) = M N/(L N + M) ≤ M/L. With M = O(1/ϵ) and L = O(1), the claimed factor O(1/ϵ) is simply the assumed ratio of task-pool size to coreset size. No property of the selected tasks—diversity, gradient matching, or submodular maximization—enters this ratio; any subset of size L gives the same factor. The sample-complexity 'reduction' is therefore encoded in the definition of S S_c and the scaling assumptions, i.e., it holds by construction rather than being a consequence of coreset selection.
-
other
[Appendix 6.3, 'Task Selection' paragraph]
"and by taking the maximum of both sides with respect to θ ∈ Θ, we obtain 1/M Σ_{j∈M} min_{i∈S} max_{θ∈Θ} ∥gj(θ) − gi(θ)∥ ≤ C e^{-1}/M + (1 − e^{-1})/M Σ_{j∈M} min_{i∈S⋆} max_{θ∈Θ} ∥gj(θ) − gi(θ)∥."
The preceding Nemhauser greedy guarantee applies to F(S) = C − Σ_j min_{i∈S} ∥g_j(θ0) − g_i(θ0)∥ evaluated at the selection point θ0. Replacing the sum over tasks by a sum of per-task max-norm terms requires an interchange of max over θ with summation and min over i; this interchange is not valid, and S is not greedy for the max-norm objective. The displayed inequality is exactly the max-norm task-selection bias bound that Theorem 1 needs, but it is obtained by assuming the desired conclusion rather than by deriving it from the coreset construction. Thus the bias term that makes the convergence guarantee non-tautological is not established; the theorem's ε-close conclusion rests on this unproved step.
full rationale
The paper's headline O(1/ε) sample-complexity reduction is largely definitional: with S S_c = (L/M)S M_c + O(M ns), M = O(1/ε), and L = O(1), the ratio S M_c/S S_c ≤ M/L = O(1/ε) follows from the assumed sizes alone. Any subset of L tasks, random or adversarial, would yield the same factor; the selection algorithm's quality does not enter the ratio. The non-tautological part of the claim is Theorem 1's N = O(1/ε) convergence when training on the selected coreset, and that part is not supported by the proof as written: Appendix 6.3 derives the max-over-θ bias bound from a θ0-greedy inequality by an invalid max/Σ interchange, so the task-selection bias in Theorem 1 is effectively assumed rather than proved. This is a correctness gap more than a self-citation chain. The self-citations to Toso et al. (2024b) and Zhan and Anderson (2024) are to published, externally grounded works and are not the mechanism that forces the result; the MAML-LQR analysis legitimately reuses prior lemmas. Overall the central sample-complexity claim reduces by construction, while the convergence proof still has independent intended content, so the circularity score is 6 rather than 8 or 10.
Assumptions & free parameters
free parameters (2)
- C (facility-location upper bound) =
C = O(epsilon) (assumed, not estimated)
- L (coreset size) =
L = O(1) in theory; 25% of 800 tasks (cartpole/hopper), 20% of 1600 tasks (walker2D), 10 of 40 tasks in LQR
assumptions (8)
- domain assumption Task rewards and gradients are locally smooth with constants beta and psi (Assumption 1).
- domain assumption Task-specific gradient norms are uniformly bounded by phi (Assumption 2).
- ad hoc to paper The constant C in (4) is set sufficiently small, C = O(epsilon) (Assumption 3).
- domain assumption Initial controller K0 in G (Assumption 4).
- domain assumption Task heterogeneity is bounded by epsilon_A, epsilon_B, epsilon_Q, epsilon_R (Assumption 5).
- domain assumption Lemma 1 from Toso et al. (2024b): ||nabla J_i(K) - nabla J_j(K)|| <= f(epsilon_het) for all stabilizing K.
- domain assumption Exact expected-reward oracle J_j(theta) is available (Remark 1).
- standard math Matrix Bernstein inequality (Tropp, 2012) and 1 - 1/e greedy submodularity guarantee (Nemhauser et al., 1978).
Cite this review
Pith. "Pith review of Coreset-Based Task Selection for Sample-Efficient Meta-Reinforcement Learning." pith.science (2026). https://pith.science/paper/DMP6OYWQ
@misc{pith2026250202332,
author = {Pith},
title = {Pith review of: Coreset-Based Task Selection for Sample-Efficient Meta-Reinforcement Learning},
year = {2026},
howpublished = {\url{https://pith.science/paper/DMP6OYWQ}},
note = {Machine review of arXiv:2502.02332}
}
abstract
We study task selection to enhance sample efficiency in model-agnostic meta-reinforcement learning (MAML-RL). Traditional meta-RL typically assumes that all available tasks are equally important, which can lead to task redundancy when they share significant similarities. To address this, we propose a coreset-based task selection approach that selects a weighted subset of tasks based on how diverse they are in gradient space, prioritizing the most informative and diverse tasks. Such task selection reduces the number of samples needed to find an $\epsilon$-close stationary solution by a factor of O(1/$\epsilon$). Consequently, it guarantees a faster adaptation to unseen tasks while focusing training on the most relevant tasks. As a case study, we incorporate task selection to MAML-LQR (Toso et al., 2024b), and prove a sample complexity reduction proportional to O(log(1/$\epsilon$)) when the task specific cost also satisfy gradient dominance. Our theoretical guarantees underscore task selection as a key component for scalable and sample-efficient meta-RL. We numerically validate this trend across multiple RL benchmark problems, illustrating the benefits of task selection beyond the LQR baseline.
Figures
Forward citations
Cited by 2 Pith papers
-
On the Gradient Domination of the LQG Problem
The LQG cost becomes gradient dominated under a history-based controller parameterization, yielding global convergence guarantees for policy gradient methods in model-based and model-free settings.
-
The Role of Diversity in In-Context Learning for Large Language Models
Diversity-aware selection of in-context examples improves performance on complex and out-of-distribution tasks, though effect sizes are often modest.
Reference graph
Works this paper leans on
-
[1]
A Moreau Envelope Approach for LQR Meta-Policy Estimation
A. Aravind, M. T. Toghani, and C. A. Uribe. A Moreau envelope approach for LQR meta-policy estimation.arXiv preprint arXiv:2403.17364,
-
[3]
(b): we first note that ∥∇ ˜J(θ) − e∇J(θ)∥ = ∥Ej,u e∇J(θ) − e∇J(θ)∥. In addition, since the tasks Tj ∈ Mand samples u ∼ Sr are drawn independently, we can use Lemma 3 to control ∥∇ ˜J(θ) − e∇J(θ)∥. Let us first denote Zl = d 2r2 Jj(θ + ul + ηinn∇ ˜Jj(θ + ul)) − Jj(θ − ul + ηinn∇ ˜Jj(θ − ul)) ul and Z = Ej,u[Zl]. ∥Zl∥ ≤d 2r |Jj(θ + ul + ηinn∇ ˜Jj(θ + ul)) ...
work page 2004
-
[8]
Convergence of Gradient-based MAML in LQR
11 N. Musavi and G. E. Dullerud. Convergence of Gradient-based MAML in LQR. arXiv preprint arXiv:2309.06588,
-
[13]
J. Shin, G. Kim, H. Lee, J. Han, and I. Yang. On task-relevant loss functions in meta-reinforcement learning and online LQR. arXiv preprint arXiv:2312.05465,
-
[14]
X. Song, W. Gao, Y . Yang, K. Choromanski, A. Pacchiano, and Y . Tang. ES-MAML: Simple Hessian-free meta learning. arXiv preprint arXiv:1910.01215,
work page Pith review arXiv 1910
-
[15]
X. Song, Y . Yang, K. Choromanski, K. Caluwaerts, W. Gao, C. Finn, and J. Tan. Rapidly adaptable legged robots via evolutionary meta-learning. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 3769–3776. IEEE,
work page 2020
-
[17]
E. Todorov, T. Erez, and Y . Tassa. Mujoco: A physics engine for model-based control. In 2012 IEEE/RSJ International Conference on Intelligent Robots and Systems, pages 5026–5033. IEEE,
work page 2012
-
[19]
L. F. Toso, H. Wang, and J. Anderson. Asynchronous Heterogeneous Linear Quadratic Regulator Design. arXiv preprint arXiv:2404.09061, 2024a. L. F. Toso, D. Zhan, J. Anderson, and H. Wang. Meta-learning linear quadratic regulators: a policy gradient MAML approach for model-free LQR. In 6th Annual Learning for Dynamics & Control Conference, pages 902–915. PM...
Show all 28 references
-
[20]
H. Wang, L. F. Toso, A. Mitra, and J. Anderson. Model-free Learning with Heterogeneous Dynamical Systems: A Federated LQR Approach. arXiv preprint arXiv:2308.11743,
-
[21]
J. X. Wang, Z. Kurth-Nelson, D. Tirumala, H. Soyer, J. Z. Leibo, R. Munos, C. Blundell, D. Kumaran, and M. Botvinick. Learning to reinforcement learn. arXiv preprint arXiv:1611.05763,
-
[23]
W. Yu, J. Tan, C. K. Liu, and G. Turk. Preparing for the unknown: Learning a universal policy with online system identification. arXiv preprint arXiv:1702.02453,
-
[24]
13 6 Appendix Roadmap: This appendix is organized as follows: First we extend our related work section and remind the reader of the model-agnostic meta-reinforcement learning problem and the matrix Bernstein inequality from Tropp (2012), where the later is crucial for controll...
2012
-
[25]
Similarly, we consider a zeroth-order estimation of the task-specific and meta-gradients
that estimate the meta-gradient through evolutionary strategy. Similarly, we consider a zeroth-order estimation of the task-specific and meta-gradients. In contrast, these works treat all the tasks equally, leading to task redundancy which we handle with a derivative-free core...
2020
-
[26]
In particular, Zhan and Anderson (2024) does not focus on RL tasks and approximates gradients using the pre-activation outputs of the last layer for classification tasks
are particularly relevant to this paper. In particular, Zhan and Anderson (2024) does not focus on RL tasks and approximates gradients using the pre-activation outputs of the last layer for classification tasks. That simplifies the problem but prevent them from deriving sample...
2024
-
[28]
In particular, by (Nemhauser et al., 1978, Section 4), we know that the value of the greedy optimization is close to the optimal as F (S) ≥ (1 − e−1)F (S ⋆)
returns a subsetS that is a suboptimal solution of the following submodular maximization S ⋆ = argmax S⊆M F (S) := C − X j∈M min i∈S ∥gj(θ0) − gi(θ0)∥ , subject to |S| ≤L, for any θ0 ∈ Θ. In particular, by (Nemhauser et al., 1978, Section 4), we know that the value of the gree...
1978
-
[1972]
S. U. Stich. Local SGD converges fast and communicates little. arXiv preprint arXiv:1805.09767,
-
[1977]
Y . Duan, J. Schulman, X. Chen, P. L. Bartlett, I. Sutskever, and P. Abbeel.RL2: Fast reinforcement learning via slow reinforcement learning. arXiv preprint arXiv:1611.02779,
-
[1982]
Y . Yang, H. Kang, and B. Mirzasoleiman. Towards sustainable learning: Coresets for data-efficient deep learning. arXiv preprint arXiv:2306.01244,
-
[2004]
Ghadirzadeh, X
A. Ghadirzadeh, X. Chen, P. Poklukar, C. Finn, M. Björkman, and D. Kragic. Bayesian meta-learning for few-shot policy adaptation across robotic platforms. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1274–1280. IEEE,
2021
-
[2012]
doi: 10.1109/ IROS.2012.6386109. 12 L. F. Toso, H. Wang, and J. Anderson. Learning personalized models with clustered system identification. In 2023 62nd IEEE Conference on Decision and Control (CDC), pages 7162–7169. IEEE,
2012
-
[2017]
A. D. Flaxman, A. T. Kalai, and H. B. McMahan. Online convex optimization in the bandit setting: gradient descent without a gradient. arXiv preprint cs/0408007,
-
[2018]
Salimans, J
T. Salimans, J. Ho, X. Chen, S. Sidor, and I. Sutskever. Evolution strategies as a scalable alternative to reinforce- ment learning. arXiv preprint arXiv:1703.03864,
-
[2019]
Pan and Q
Y . Pan and Q. Zhu. Model-agnostic zeroth-order policy optimization for meta-learning of ergodic linear quadratic regulators. arXiv preprint arXiv:2405.17370,
-
[2020]
Molybog and J
I. Molybog and J. Lavaei. When does maml objective have benign landscape? In 2021 IEEE Conference on Control Technology and Applications (CCTA), pages 220–227. IEEE,
2021
-
[2021]
B. Lee, A. Rantzer, and N. Matni. Nonasymptotic regret analysis of adaptive linear quadratic control with model misspecification. In 6th Annual Learning for Dynamics & Control Conference, pages 980–992. PMLR, 2024a. B. D. Lee, L. F. Toso, T. T. Zhang, J. Anderson, and N. Matni...
1907 arXiv
-
[2022]
Rothfuss, D
J. Rothfuss, D. Lee, I. Clavera, T. Asfour, and P. Abbeel. Promp: Proximal meta-policy search. arXiv preprint arXiv:1810.06784,
-
[2023]
Nagabandi, I
A. Nagabandi, I. Clavera, S. Liu, R. S. Fearing, P. Abbeel, S. Levine, and C. Finn. Learning to adapt in dynamic, real-world environments through meta-reinforcement learning. arXiv preprint arXiv:1803.11347,
-
[2024]
Arndt, M
K. Arndt, M. Hazara, A. Ghadirzadeh, and V . Kyrki. Meta reinforcement learning for sim-to-real domain adaptation. In 2020 IEEE international conference on robotics and automation (ICRA) , pages 2725–2731. IEEE,
2020
Reviewed August 9, 2026 · model on record in the stance chip above.
Discussion (0). Continue with ORCID to comment.