Shie Mannor

Eli A. Meirom, Haggai Maron, Shie Mannor et al. · nvidia

Quantum Computing (QC) stands to revolutionize computing, but is currently still limited. To develop and test quantum algorithms today, quantum circuits are often simulated on classical computers. Simulating a complex quantum circuit requires computing the contraction of a large network of tensors. The order (path) of contraction can have a drastic effect on the computing cost, but finding an efficient order is a challenging combinatorial optimization problem. We propose a Reinforcement Learning (RL) approach combined with Graph Neural Networks (GNN) to address the contraction ordering problem. The problem is extremely challenging due to the huge search space, the heavy-tailed reward distribution, and the challenging credit assignment. We show how a carefully implemented RL-agent that uses a GNN as the basic policy construct can address these challenges and obtain significant improvements over state-of-the-art techniques in three varieties of circuits, including the largest scale networks used in contemporary QC.

Ido Greenberg, Shie Mannor, Gal Chechik et al. · nvidia

A major challenge of reinforcement learning (RL) in real-world applications is the variation between environments, tasks or clients. Meta-RL (MRL) addresses this issue by learning a meta-policy that adapts to new tasks. Standard MRL methods optimize the average return over tasks, but often suffer from poor results in tasks of high risk or difficulty. This limits system reliability since test tasks are not known in advance. In this work, we define a robust MRL objective with a controlled robustness level. Optimization of analogous robust objectives in RL is known to lead to both *biased gradients* and *data inefficiency*. We prove that the gradient bias disappears in our proposed MRL framework. The data inefficiency is addressed via the novel Robust Meta RL algorithm (RoML). RoML is a meta-algorithm that generates a robust version of any given MRL algorithm, by identifying and over-sampling harder tasks throughout training. We demonstrate that RoML achieves robust returns on multiple navigation and continuous control benchmarks.

Benjamin Fuhrer, Yuval Shpigelman, Chen Tessler et al. · nvidia

As communication protocols evolve, datacenter network utilization increases. As a result, congestion is more frequent, causing higher latency and packet loss. Combined with the increasing complexity of workloads, manual design of congestion control (CC) algorithms becomes extremely difficult. This calls for the development of AI approaches to replace the human effort. Unfortunately, it is currently not possible to deploy AI models on network devices due to their limited computational capabilities. Here, we offer a solution to this problem by building a computationally-light solution based on a recent reinforcement learning CC algorithm [arXiv:2207.02295]. We reduce the inference time of RL-CC by x500 by distilling its complex neural network into decision trees. This transformation enables real-time inference within the $μ$-sec decision-time requirement, with a negligible effect on quality. We deploy the transformed policy on NVIDIA NICs in a live cluster. Compared to popular CC algorithms used in production, RL-CC is the only method that performs well on all benchmarks tested over a large range of number of flows. It balances multiple metrics simultaneously: bandwidth, latency, and packet drops. These results suggest that data-driven methods for CC are feasible, challenging the prior belief that handcrafted heuristics are necessary to achieve optimal performance.

Guy Tennenholtz, Nadav Merlis, Lior Shani et al. · nvidia

We present the problem of reinforcement learning with exogenous termination. We define the Termination Markov Decision Process (TerMDP), an extension of the MDP framework, in which episodes may be interrupted by an external non-Markovian observer. This formulation accounts for numerous real-world situations, such as a human interrupting an autonomous driving agent for reasons of discomfort. We learn the parameters of the TerMDP and leverage the structure of the estimation problem to provide state-wise confidence bounds. We use these to construct a provably-efficient algorithm, which accounts for termination, and bound its regret. Motivated by our theoretical analysis, we design and implement a scalable approach, which combines optimism (w.r.t. termination) and a dynamic discount factor, incorporating the termination probability. We deploy our method on high-dimensional driving and MinAtar benchmarks. Additionally, we test our approach on human data in a driving setting. Our results demonstrate fast convergence and significant improvement over various baseline approaches.

Peter Karkus, Boris Ivanovic, Shie Mannor et al.

Autonomous vehicle (AV) stacks are typically built in a modular fashion, with explicit components performing detection, tracking, prediction, planning, control, etc. While modularity improves reusability, interpretability, and generalizability, it also suffers from compounding errors, information bottlenecks, and integration challenges. To overcome these challenges, a prominent approach is to convert the AV stack into an end-to-end neural network and train it with data. While such approaches have achieved impressive results, they typically lack interpretability and reusability, and they eschew principled analytical components, such as planning and control, in favor of deep neural networks. To enable the joint optimization of AV stacks while retaining modularity, we present DiffStack, a differentiable and modular stack for prediction, planning, and control. Crucially, our model-based planning and control algorithms leverage recent advancements in differentiable optimization to produce gradients, enabling optimization of upstream components, such as prediction, via backpropagation through planning and control. Our results on the nuScenes dataset indicate that end-to-end training with DiffStack yields substantial improvements in open-loop and closed-loop planning metrics by, e.g., learning to make fewer prediction errors that would affect planning. Beyond these immediate benefits, DiffStack opens up new opportunities for fully data-driven yet modular and interpretable AV architectures. Project website: https://sites.google.com/view/diffstack

Gal Dalal, Assaf Hallak, Gugan Thoppe et al. · nvidia

Policy gradient methods are notorious for having a large variance and high sample complexity. To mitigate this, we introduce SoftTreeMax -- a generalization of softmax that employs planning. In SoftTreeMax, we extend the traditional logits with the multi-step discounted cumulative reward, topped with the logits of future states. We analyze SoftTreeMax and explain how tree expansion helps to reduce its gradient variance. We prove that the variance depends on the chosen tree-expansion policy. Specifically, we show that the closer the induced transitions are to being state-independent, the stronger the variance decay. With approximate forward models, we prove that the resulting gradient bias diminishes with the approximation error while retaining the same variance reduction. Ours is the first result to bound the gradient bias for an approximate model. In a practical implementation of SoftTreeMax, we utilize a parallel GPU-based simulator for fast and efficient tree expansion. Using this implementation in Atari, we show that SoftTreeMax reduces the gradient variance by three orders of magnitude. This leads to better sample complexity and improved performance compared to distributed PPO.

Ido Greenberg, Yinlam Chow, Mohammad Ghavamzadeh et al.

In risk-averse reinforcement learning (RL), the goal is to optimize some risk measure of the returns. A risk measure often focuses on the worst returns out of the agent's experience. As a result, standard methods for risk-averse RL often ignore high-return strategies. We prove that under certain conditions this inevitably leads to a local-optimum barrier, and propose a soft risk mechanism to bypass it. We also devise a novel Cross Entropy module for risk sampling, which (1) preserves risk aversion despite the soft risk; (2) independently improves sample efficiency. By separating the risk aversion of the sampler and the optimizer, we can sample episodes with poor conditions, yet optimize with respect to successful strategies. We combine these two concepts in CeSoR - Cross-entropy Soft-Risk optimization algorithm - which can be applied on top of any risk-averse policy gradient (PG) method. We demonstrate improved risk aversion in maze navigation, autonomous driving, and resource allocation benchmarks, including in scenarios where standard risk-averse PG completely fails.

Navdeep Kumar, Esther Derman, Matthieu Geist et al.

Policy gradient methods have become a standard for training reinforcement learning agents in a scalable and efficient manner. However, they do not account for transition uncertainty, whereas learning robust policies can be computationally expensive. In this paper, we introduce robust policy gradient (RPG), a policy-based method that efficiently solves rectangular robust Markov decision processes (MDPs). We provide a closed-form expression for the worst occupation measure. Incidentally, we find that the worst kernel is a rank-one perturbation of the nominal. Combining the worst occupation measure with a robust Q-value estimation yields an explicit form of the robust gradient. Our resulting RPG can be estimated from data with the same time complexity as its non-robust equivalent. Hence, it relieves the computational burden of convex optimization problems required for training robust policies by current policy gradient approaches.

Gal Dalal, Assaf Hallak, Shie Mannor et al. · nvidia

Policy-gradient methods are widely used for learning control policies. They can be easily distributed to multiple workers and reach state-of-the-art results in many domains. Unfortunately, they exhibit large variance and subsequently suffer from high-sample complexity since they aggregate gradients over entire trajectories. At the other extreme, planning methods, like tree search, optimize the policy using single-step transitions that consider future lookahead. These approaches have been mainly considered for value-based algorithms. Planning-based algorithms require a forward model and are computationally intensive at each step, but are more sample efficient. In this work, we introduce SoftTreeMax, the first approach that integrates tree-search into policy gradient. Traditionally, gradients are computed for single state-action pairs. Instead, our tree-based policy structure leverages all gradients at the tree leaves in each environment step. This allows us to reduce the variance of gradients by three orders of magnitude and to benefit from better sample complexity compared with standard policy gradient. On Atari, SoftTreeMax demonstrates up to 5x better performance in faster run-time compared with distributed PPO.

Navdeep Kumar, Kfir Levy, Kaixin Wang et al.

Robust Markov decision processes (MDPs) provide a general framework to model decision problems where the system dynamics are changing or only partially known. Efficient methods for some \texttt{sa}-rectangular robust MDPs exist, using its equivalence with reward regularized MDPs, generalizable to online settings. In comparison to \texttt{sa}-rectangular robust MDPs, \texttt{s}-rectangular robust MDPs are less restrictive but much more difficult to deal with. Interestingly, recent works have established the equivalence between \texttt{s}-rectangular robust MDPs and policy regularized MDPs. But we don't have a clear understanding to exploit this equivalence, to do policy improvement steps to get the optimal value function or policy. We don't have a clear understanding of greedy/optimal policy except it can be stochastic. There exist no methods that can naturally be generalized to model-free settings. We show a clear and explicit equivalence between \texttt{s}-rectangular $L_p$ robust MDPs and policy regularized MDPs that resemble very much policy entropy regularized MDPs widely used in practice. Further, we dig into the policy improvement step and concretely derive optimal robust Bellman operators for \texttt{s}-rectangular $L_p$ robust MDPs. We find that the greedy/optimal policies in \texttt{s}-rectangular $L_p$ robust MDPs are threshold policies that play top $k$ actions whose $Q$ value is greater than some threshold (value), proportional to the $(p-1)$th power of its advantage. In addition, we show time complexity of (\texttt{sa} and \texttt{s}-rectangular) $L_p$ robust MDPs is the same as non-robust MDPs up to some log factors. Our work greatly extends the existing understanding of \texttt{s}-rectangular robust MDPs and naturally generalizable to online settings.

Kaixin Wang, Uri Gadot, Navdeep Kumar et al.

Robust Markov Decision Processes (RMDPs) provide a framework for sequential decision-making that is robust to perturbations on the transition kernel. However, current RMDP methods are often limited to small-scale problems, hindering their use in high-dimensional domains. To bridge this gap, we present EWoK, a novel online approach to solve RMDP that Estimates the Worst transition Kernel to learn robust policies. Unlike previous works that regularize the policy or value updates, EWoK achieves robustness by simulating the worst scenarios for the agent while retaining complete flexibility in the learning process. Notably, EWoK can be applied on top of any off-the-shelf {\em non-robust} RL algorithm, enabling easy scaling to high-dimensional domains. Our experiments, spanning from simple Cartpole to high-dimensional DeepMind Control Suite environments, demonstrate the effectiveness and applicability of the EWoK paradigm as a practical method for learning robust policies.

Shirli Di Castro Shashua, Shie Mannor, Dotan Di-Castro

Replay buffers are a key component in many reinforcement learning schemes. Yet, their theoretical properties are not fully understood. In this paper we analyze a system where a stochastic process X is pushed into a replay buffer and then randomly sampled to generate a stochastic process Y from the replay buffer. We provide an analysis of the properties of the sampled process such as stationarity, Markovity and autocorrelation in terms of the properties of the original process. Our theoretical analysis sheds light on why replay buffer may be a good de-correlator. Our analysis provides theoretical tools for proving the convergence of replay buffer based algorithms which are prevalent in reinforcement learning schemes.

Navdeep Kumar, Kfir Levy, Kaixin Wang et al.

We present an efficient robust value iteration for \texttt{s}-rectangular robust Markov Decision Processes (MDPs) with a time complexity comparable to standard (non-robust) MDPs which is significantly faster than any existing method. We do so by deriving the optimal robust Bellman operator in concrete forms using our $L_p$ water filling lemma. We unveil the exact form of the optimal policies, which turn out to be novel threshold policies with the probability of playing an action proportional to its advantage.

Gal Dalal, Elad Gilboa, Shie Mannor

Asset management attempts to keep the power system in working conditions. It requires much coordination between multiple entities and long term planning often months in advance. In this work we introduce a mid-term asset management formulation as a stochastic optimization problem, that includes three hierarchical layers of decision making, namely the mid-term, short-term and real-time. We devise a tractable scenario approximation technique for efficiently assessing the complex implications a maintenance schedule inflicts on a power system. This is done using efficient Monte-Carlo simulations that trade-off between accuracy and tractability. We then present our implementation of a distributed scenario-based optimization algorithm for solving our formulation, and use an updated PJM 5-bus system to show a solution that is cheaper than other maintenance heuristics that are likely to be considered by TSOs.

Shie Mannor, Aviv Tamar

Reinforcement learning (RL) has demonstrated great potential, but is currently full of overhyping and pipe dreams. We point to some difficulties with current research which we feel are endemic to the direction taken by the community. To us, the current direction is not likely to lead to "deployable" RL: RL that works in practice and can work in practical situations yet still is economically viable. We also propose a potential fix to some of the difficulties of the field.

Navdeep Kumar, Kaixin Wang, Kfir Levy et al.

In Reinforcement Learning (RL), the goal of agents is to discover an optimal policy that maximizes the expected cumulative rewards. This objective may also be viewed as finding a policy that optimizes a linear function of its state-action occupancy measure, hereafter referred as Linear RL. However, many supervised and unsupervised RL problems are not covered in the Linear RL framework, such as apprenticeship learning, pure exploration and variational intrinsic control, where the objectives are non-linear functions of the occupancy measures. RL with non-linear utilities looks unwieldy, as methods like Bellman equation, value iteration, policy gradient, dynamic programming that had tremendous success in Linear RL, fail to trivially generalize. In this paper, we derive the policy gradient theorem for RL with general utilities. The policy gradient theorem proves to be a cornerstone in Linear RL due to its elegance and ease of implementability. Our policy gradient theorem for RL with general utilities shares the same elegance and ease of implementability. Based on the policy gradient theorem derived, we also present a simple sample-based algorithm. We believe our results will be of interest to the community and offer inspiration to future works in this generalized setting.

Jeongyeol Kwon, Yonathan Efroni, Constantine Caramanis et al.

We consider episodic reinforcement learning in reward-mixing Markov decision processes (RMMDPs): at the beginning of every episode nature randomly picks a latent reward model among $M$ candidates and an agent interacts with the MDP throughout the episode for $H$ time steps. Our goal is to learn a near-optimal policy that nearly maximizes the $H$ time-step cumulative rewards in such a model. Previous work established an upper bound for RMMDPs for $M=2$. In this work, we resolve several open questions remained for the RMMDP model. For an arbitrary $M\ge2$, we provide a sample-efficient algorithm--$\texttt{EM}^2$--that outputs an $ε$-optimal policy using $\tilde{O} \left(ε^{-2} \cdot S^d A^d \cdot \texttt{poly}(H, Z)^d \right)$ episodes, where $S, A$ are the number of states and actions respectively, $H$ is the time-horizon, $Z$ is the support size of reward distributions and $d=\min(2M-1,H)$. Our technique is a higher-order extension of the method-of-moments based approach, nevertheless, the design and analysis of the \algname algorithm requires several new ideas beyond existing techniques. We also provide a lower bound of $(SA)^{Ω(\sqrt{M})} / ε^{2}$ for a general instance of RMMDP, supporting that super-polynomial sample complexity in $M$ is necessary.

Mohammadi Zaki, Avinash Mohan, Aditya Gopalan et al.

We consider an improper reinforcement learning setting where a learner is given $M$ base controllers for an unknown Markov decision process, and wishes to combine them optimally to produce a potentially new controller that can outperform each of the base ones. This can be useful in tuning across controllers, learnt possibly in mismatched or simulated environments, to obtain a good controller for a given target environment with relatively few trials. Towards this, we propose two algorithms: (1) a Policy Gradient-based approach; and (2) an algorithm that can switch between a simple Actor-Critic (AC) based scheme and a Natural Actor-Critic (NAC) scheme depending on the available information. Both algorithms operate over a class of improper mixtures of the given controllers. For the first case, we derive convergence rate guarantees assuming access to a gradient oracle. For the AC-based approach we provide convergence rate guarantees to a stationary point in the basic AC case and to a global optimum in the NAC case. Numerical results on (i) the standard control theoretic benchmark of stabilizing an cartpole; and (ii) a constrained queueing task show that our improper policy optimization algorithm can stabilize the system even when the base policies at its disposal are unstable.

Jeongyeol Kwon, Yonathan Efroni, Constantine Caramanis et al.

We consider a multi-armed bandit problem with $M$ latent contexts, where an agent interacts with the environment for an episode of $H$ time steps. Depending on the length of the episode, the learner may not be able to estimate accurately the latent context. The resulting partial observation of the environment makes the learning task significantly more challenging. Without any additional structural assumptions, existing techniques to tackle partially observed settings imply the decision maker can learn a near-optimal policy with $O(A)^H$ episodes, but do not promise more. In this work, we show that learning with {\em polynomial} samples in $A$ is possible. We achieve this by using techniques from experiment design. Then, through a method-of-moments approach, we design a procedure that provably learns a near-optimal policy with $O(\texttt{poly}(A) + \texttt{poly}(M,H)^{\min(M,H)})$ interactions. In practice, we show that we can formulate the moment-matching via maximum likelihood estimation. In our experiments, this significantly outperforms the worst-case guarantees, as well as existing practical methods.

Esther Derman, Yevgeniy Men, Matthieu Geist et al.

Robust Markov decision processes (MDPs) aim to handle changing or partially known system dynamics. To solve them, one typically resorts to robust optimization methods. However, this significantly increases computational complexity and limits scalability in both learning and planning. On the other hand, regularized MDPs show more stability in policy learning without impairing time complexity. Yet, they generally do not encompass uncertainty in the model dynamics. In this work, we aim to learn robust MDPs using regularization. We first show that regularized MDPs are a particular instance of robust MDPs with uncertain reward. We thus establish that policy iteration on reward-robust MDPs can have the same time complexity as on regularized MDPs. We further extend this relationship to MDPs with uncertain transitions: this leads to a regularization term with an additional dependence on the value function. We then generalize regularized MDPs to twice regularized MDPs ($\text{R}^2$ MDPs), i.e., MDPs with $\textit{both}$ value and policy regularization. The corresponding Bellman operators enable us to derive planning and learning schemes with convergence and generalization guarantees, thus reducing robustness to regularization. We numerically show this two-fold advantage on tabular and physical domains, highlighting the fact that $\text{R}^2$ preserves its efficacy in continuous environments.

Jeongyeol Kwon, Yonathan Efroni, Shie Mannor et al.

In many interactive decision-making settings, there is latent and unobserved information that remains fixed. Consider, for example, a dialogue system, where complete information about a user, such as the user's preferences, is not given. In such an environment, the latent information remains fixed throughout each episode, since the identity of the user does not change during an interaction. This type of environment can be modeled as a Latent Markov Decision Process (LMDP), a special instance of Partially Observed Markov Decision Processes (POMDPs). Previous work established exponential lower bounds in the number of latent contexts for the LMDP class. This puts forward a question: under which natural assumptions a near-optimal policy of an LMDP can be efficiently learned? In this work, we study the class of LMDPs with {\em prospective side information}, when an agent receives additional, weakly revealing, information on the latent context at the beginning of each episode. We show that, surprisingly, this problem is not captured by contemporary settings and algorithms designed for partially observed environments. We then establish that any sample efficient algorithm must suffer at least $Ω(K^{2/3})$-regret, as opposed to standard $Ω(\sqrt{K})$ lower bounds, and design an algorithm with a matching upper bound.

Guy Lutsker, Gal Sapir, Smadar Shilo et al.

Recent advances in SSL enabled novel medical AI models, known as foundation models, offer great potential for better characterizing health from diverse biomedical data. CGM provides rich, temporal data on glycemic patterns, but its full potential for predicting broader health outcomes remains underutilized. Here, we present GluFormer, a generative foundation model for CGM data that learns nuanced glycemic patterns and translates them into predictive representations of metabolic health. Trained on over 10 million CGM measurements from 10,812 adults, primarily without diabetes, GluFormer uses autoregressive token prediction to capture longitudinal glucose dynamics. We show that GluFormer generalizes to 19 external cohorts (n=6,044) spanning different ethnicities and ages, 5 countries, 8 CGM devices, and diverse pathophysiological states. GluFormers representations exceed the performance of current CGM metrics, such as the Glucose Management Indicator (GMI), for forecasting clinical measures. In a longitudinal study of 580 adults with CGM data and 12-year follow-up, GluFormer identifies individuals at elevated risk of developing diabetes more effectively than blood HbA1C%, capturing 66% of all new-onset diabetes diagnoses in the top quartile versus 7% in the bottom quartile. Similarly, 69% of cardiovascular-death events occurred in the top quartile with none in the bottom quartile, demonstrating powerful risk stratification beyond traditional glycemic metrics. We also show that CGM representations from pre-intervention periods in Randomized Clinical Trials outperform other methods in predicting primary and secondary outcomes. When integrating dietary data into GluFormer, we show that the multi-modal version of the model can accurately generate CGM data based on dietary intake data, simulate outcomes of dietary interventions, and predict individual responses to specific foods.

Stav Belogolovsky, Ido Greenberg, Danny Eytan et al.

Dosing models often use differential equations to model biological dynamics. Neural differential equations in particular can learn to predict the derivative of a process, which permits predictions at irregular points of time. However, this temporal flexibility often comes with a high sensitivity to noise, whereas medical problems often present high noise and limited data. Moreover, medical dosing models must generalize reliably over individual patients and changing treatment policies. To address these challenges, we introduce the Neural Eigen Stochastic Differential Equation algorithm (NESDE). NESDE provides individualized modeling (using a hypernetwork over patient-level parameters); generalization to new treatment policies (using decoupled control); tunable expressiveness according to the noise level (using piecewise linearity); and fast, continuous, closed-form prediction (using spectral representation). We demonstrate the robustness of NESDE in both synthetic and real medical problems, and use the learned dynamics to publish simulated medical gym environments.

Arman C. Kizilkale, Shie Mannor

We analyze the efficiency of markets with friction, particularly power markets. We model the market as a dynamic system with $(d_t;\,t\geq 0)$ the demand process and $(s_t;\,t\geq 0)$ the supply process. Using stochastic differential equations to model the dynamics with friction, we investigate the efficiency of the market under an integrated expected undiscounted cost function solving the optimal control problem. Then, we extend the setup to a game theoretic model where multiple suppliers and consumers interact continuously by setting prices in a dynamic market with friction. We investigate the equilibrium, and analyze the efficiency of the market under an integrated expected social cost function. We provide an intriguing efficiency-volatility no-free-lunch trade-off theorem.

Ido Greenberg, Netanel Yannay, Shie Mannor

In non-linear filtering, it is traditional to compare non-linear architectures such as neural networks to the standard linear Kalman Filter (KF). We observe that this mixes the evaluation of two separate components: the non-linear architecture, and the parameters optimization method. In particular, the non-linear model is often optimized, whereas the reference KF model is not. We argue that both should be optimized similarly, and to that end present the Optimized KF (OKF). We demonstrate that the KF may become competitive to neural models - if optimized using OKF. This implies that experimental conclusions of certain previous studies were derived from a flawed process. The advantage of OKF over the standard KF is further studied theoretically and empirically, in a variety of problems. Conveniently, OKF can replace the KF in real-world systems by merely updating the parameters.

Navdeep Kumar, Tehila Dahan, Lior Cohen et al.

We establish an optimal sample complexity of $O(ε^{-2})$ for obtaining an $ε$-optimal global policy using a single-timescale actor-critic (AC) algorithm in infinite-horizon discounted Markov decision processes (MDPs) with finite state-action spaces, improving upon the prior state of the art of $O(ε^{-3})$. Our approach applies STORM (STOchastic Recursive Momentum) to reduce variance in the critic updates. However, because samples are drawn from a nonstationary occupancy measure induced by the evolving policy, variance reduction via STORM alone is insufficient. To address this challenge, we maintain a buffer of small fraction of recent samples and uniformly sample from it for each critic update. Importantly, these mechanisms are compatible with existing deep learning architectures and require only minor modifications, without compromising practical applicability.

Elad Sharony, Tom Jurgenson, Orr Krupnik et al.

Recent advances in Large Language Models (LLMs) and Vision-Language Models (VLMs) have enabled powerful semantic and multimodal reasoning capabilities, creating new opportunities to enhance sample efficiency, high-level planning, and interpretability in reinforcement learning (RL). While prior work has integrated LLMs and VLMs into various components of RL, the replay buffer, a core component for storing and reusing experiences, remains unexplored. We propose addressing this gap by leveraging VLMs to guide the prioritization of experiences in the replay buffer. Our key idea is to use a frozen, pre-trained VLM (requiring no fine-tuning) as an automated evaluator to identify and prioritize promising sub-trajectories from the agent's experiences. Across scenarios, including game-playing and robotics, spanning both discrete and continuous domains, agents trained with our proposed prioritization method achieve 11-52% higher average success rates and improve sample efficiency by 19-45% compared to previous approaches. https://esharony.me/projects/vlm-rb/

Mark Kozdoba, Binyamin Perets, Shie Mannor

We propose a new approach to non-parametric density estimation that is based on regularizing a Sobolev norm of the density. This method is statistically consistent, and makes the inductive bias of the model clear and interpretable. While there is no closed analytic form for the associated kernel, we show that one can approximate it using sampling. The optimization problem needed to determine the density is non-convex, and standard gradient methods do not perform well. However, we show that with an appropriate initialization and using natural gradients, one can obtain well performing solutions. Finally, while the approach provides pre-densities (i.e. not necessarily integrating to 1), which prevents the use of log-likelihood for cross validation, we show that one can instead adapt Fisher divergence based score matching methods for this task. We evaluate the resulting method on the comprehensive recent anomaly detection benchmark suite, ADBench, and find that it ranks second best, among more than 15 algorithms.

Binyamin Perets, Mark Kozdoba, Shie Mannor

The Hidden Markov Model (HMM) is one of the most widely used statistical models for sequential data analysis. One of the key reasons for this versatility is the ability of HMM to deal with missing data. However, standard HMM learning algorithms rely crucially on the assumption that the positions of the missing observations \emph{within the observation sequence} are known. In the natural sciences, where this assumption is often violated, special variants of HMM, commonly known as Silent-state HMMs (SHMMs), are used. Despite their widespread use, these algorithms strongly rely on specific structural assumptions of the underlying chain, such as acyclicity, thus limiting the applicability of these methods. Moreover, even in the acyclic case, it has been shown that these methods can lead to poor reconstruction. In this paper we consider the general problem of learning an HMM from data with unknown missing observation locations. We provide reconstruction algorithms that do not require any assumptions about the structure of the underlying chain, and can also be used with limited prior knowledge, unlike SHMM. We evaluate and compare the algorithms in a variety of scenarios, measuring their reconstruction precision, and robustness under model miss-specification. Notably, we show that under proper specifications one can reconstruct the process dynamics as well as if the missing observations positions were known.

Lior Cohen, Kaixin Wang, Bingyi Kang et al.

Motivated by the success of Transformers when applied to sequences of discrete symbols, token-based world models (TBWMs) were recently proposed as sample-efficient methods. In TBWMs, the world model consumes agent experience as a language-like sequence of tokens, where each observation constitutes a sub-sequence. However, during imagination, the sequential token-by-token generation of next observations results in a severe bottleneck, leading to long training times, poor GPU utilization, and limited representations. To resolve this bottleneck, we devise a novel Parallel Observation Prediction (POP) mechanism. POP augments a Retentive Network (RetNet) with a novel forward mode tailored to our reinforcement learning setting. We incorporate POP in a novel TBWM agent named REM (Retentive Environment Model), showcasing a 15.4x faster imagination compared to prior TBWMs. REM attains superhuman performance on 12 out of 26 games of the Atari 100K benchmark, while training in less than 12 hours. Our code is available at \url{https://github.com/leor-c/REM}.

Binyamin Perets, Shie Mannor

Large-scale hypothesis testing is central to modern science, where controlling the False Discovery Rate (FDR) has become the standard approach to managing false positives across many simultaneous tests. Hypotheses rarely exist in isolation; they often exhibit structure through proximity, connectivity, or hierarchy. This structure represents both a challenge and an opportunity: while classical methods treat these dependencies as obstacles requiring conservative correction, leveraging them can substantially increase discovery power. Here, we reframe structured FDR control as a regularized learning problem. By optimizing within a suitable Reproducing Kernel Hilbert Space (RKHS), we introduce a framework that unifies continuous domains, graphs, and hierarchies under a single algorithm through kernel choice alone. This formulation enables smooth solutions in place of the piecewise-constant fits of prior methods, principled likelihood-based hyperparameter selection rather than heuristic tuning, and inference at unobserved locations which in turn supports sample-efficient experimental design. Building on this estimator, we provide two decision rules which we prove to control the FDR. We validate our method on two sources: spatial locations derived from high-dimensional real-world datasets, and a differential gene expression task utilizing protein-protein interaction graphs.

Mark Kozdoba, Binyamin Perets, Shie Mannor

There is a well known intrinsic trade-off between the fairness of a representation and the performance of classifiers derived from the representation. Due to the complexity of optimisation algorithms in most modern representation learning approaches, for a given method it may be non-trivial to decide whether the obtained fairness-performance curve of the method is optimal, i.e., whether it is close to the true Pareto front for these quantities for the underlying data distribution. In this paper we propose a new method to compute the optimal Pareto front, which does not require the training of complex representation models. We show that optimal fair representations possess several useful structural properties, and that these properties enable a reduction of the computation of the Pareto Front to a compact discrete problem. We then also show that these compact approximating problems can be efficiently solved via off-the shelf concave-convex programming methods. Since our approach is independent of the specific model of representations, it may be used as the benchmark to which representation learning algorithms may be compared. We experimentally evaluate the approach on a number of real world benchmark datasets.

David Valensi, Esther Derman, Shie Mannor et al.

The standard formulation of Markov decision processes (MDPs) assumes that the agent's decisions are executed immediately. However, in numerous realistic applications such as robotics or healthcare, actions are performed with a delay whose value can even be stochastic. In this work, we introduce stochastic delayed execution MDPs, a new formalism addressing random delays without resorting to state augmentation. We show that given observed delay values, it is sufficient to perform a policy search in the class of Markov policies in order to reach optimal performance, thus extending the deterministic fixed delay case. Armed with this insight, we devise DEZ, a model-based algorithm that optimizes over the class of Markov policies. DEZ leverages Monte-Carlo tree search similar to its non-delayed variant EfficientZero to accurately infer future states from the action queue. Thus, it handles delayed execution while preserving the sample efficiency of EfficientZero. Through a series of experiments on the Atari suite, we demonstrate that although the previous baseline outperforms the naive method in scenarios with constant delay, it underperforms in the face of stochastic delays. In contrast, our approach significantly outperforms the baselines, for both constant and stochastic delays. The code is available at http://github.com/davidva1/Delayed-EZ .

Elad Sharony, Heng Yang, Tong Che et al.

Sequentially solving similar optimization problems under strict runtime constraints is essential for many applications, such as robot control, autonomous driving, and portfolio management. The performance of local optimization methods in these settings is sensitive to the initial solution: poor initialization can lead to slow convergence or suboptimal solutions. To address this challenge, we propose learning to predict \emph{multiple} diverse initial solutions given parameters that define the problem instance. We introduce two strategies for utilizing multiple initial solutions: (i) a single-optimizer approach, where the most promising initial solution is chosen using a selection function, and (ii) a multiple-optimizers approach, where several optimizers, potentially run in parallel, are each initialized with a different solution, with the best solution chosen afterward. Notably, by including a default initialization among predicted ones, the cost of the final output is guaranteed to be equal or lower than with the default initialization. We validate our method on three optimal control benchmark tasks: cart-pole, reacher, and autonomous driving, using different optimizers: DDP, MPPI, and iLQR. We find significant and consistent improvement with our method across all evaluation settings and demonstrate that it efficiently scales with the number of initial solutions required. The code is available at MISO (https://github.com/EladSharony/miso).

Itai Shufaro, Gal Benor, Shie Mannor

Hybrid mechanistic models, physical priors with learned residuals, promise to reduce the data required for good decisions, but have no computable criterion to test this. We characterize the value of mechanistic priors in sequential decision-making within both asymptotic and burn-in regimes. To formalize this, we introduce the mechanistic information of a model -- the mutual information between the model's recommended policy $\hatπ$ and the true optimal policy $π^*$ -- quantified via an occupancy-weighted bias $B_μ$. In the asymptotic regime (large $N$), matched bounds reveal that Bayesian regret scales with the residual entropy $H_{\mathrm{mech}}$, delivering a theoretical sample complexity reduction of $H(μ)/H_{\mathrm{mech}}$ compared to an uninformed baseline. Furthermore, we provide a model certificate to determine empirical sample efficiency. Complementarily, in the clinically relevant burn-in regime (small $N$), we establish a lower bound on the penalty incurred by confidently wrong priors. We demonstrate both the asymptotic and burn-in bounds across 5-fluorouracil (5-FU) dosing simulations motivated by published FOLFOX pharmacokinetic data, where a hybrid prior yields large sample-efficiency gains in the burn-in regime. Finally, we contrast these grounded models with LLM priors, demonstrating that LLMs can suffer severe losses in mechanistic information, thereby motivating the exclusive use of physically-grounded priors for safety-critical applications.

Lior Cohen, Kaixin Wang, Bingyi Kang et al.

World model (WM) agents enable sample-efficient reinforcement learning by learning policies entirely from simulated experience. However, existing token-based world models (TBWMs) are limited to visual inputs and discrete actions, restricting their adoption and applicability. Moreover, although both intrinsic motivation and prioritized WM replay have shown promise in improving WM performance and generalization, they remain underexplored in this setting, particularly in combination. We introduce Simulus, a highly modular TBWM agent that integrates (1) a modular multi-modality tokenization framework, (2) intrinsic motivation, (3) prioritized WM replay, and (4) regression-as-classification for reward and return prediction. Simulus achieves state-of-the-art sample efficiency for planning-free WMs across three diverse benchmarks. Ablation studies reveal the individual contribution of each component while highlighting their synergy. Our code and model weights are publicly available at https://github.com/leor-c/Simulus.

Assaf Hallak, Gal Dalal, Steven Dalton et al.

Tree Search (TS) is crucial to some of the most influential successes in reinforcement learning. Here, we tackle two major challenges with TS that limit its usability: \textit{distribution shift} and \textit{scalability}. We first discover and analyze a counter-intuitive phenomenon: action selection through TS and a pre-trained value function often leads to lower performance compared to the original pre-trained agent, even when having access to the exact state and reward in future steps. We show this is due to a distribution shift to areas where value estimates are highly inaccurate and analyze this effect using Extreme Value theory. To overcome this problem, we introduce a novel off-policy correction term that accounts for the mismatch between the pre-trained value and its corresponding TS policy by penalizing under-sampled trajectories. We prove that our correction eliminates the above mismatch and bound the probability of sub-optimal action selection. Our correction significantly improves pre-trained Rainbow agents without any further training, often more than doubling their scores on Atari games. Next, we address the scalability issue given by the computational complexity of exhaustive TS that scales exponentially with the tree depth. We introduce Batch-BFS: a GPU breadth-first search that advances all nodes in each depth of the tree simultaneously. Batch-BFS reduces runtime by two orders of magnitude and, beyond inference, enables also training with TS of depths that were not feasible before. We train DQN agents from scratch using TS and show improvement in several Atari games compared to both the original DQN and the more advanced Rainbow. The code for BCTS can be found in \url{https://github.com/NVlabs/bcts}.

Esther Derman, Gal Dalal, Shie Mannor

The standard Markov Decision Process (MDP) formulation hinges on the assumption that an action is executed immediately after it was chosen. However, assuming it is often unrealistic and can lead to catastrophic failures in applications such as robotic manipulation, cloud computing, and finance. We introduce a framework for learning and planning in MDPs where the decision-maker commits actions that are executed with a delay of $m$ steps. The brute-force state augmentation baseline where the state is concatenated to the last $m$ committed actions suffers from an exponential complexity in $m$, as we show for policy iteration. We then prove that with execution delay, deterministic Markov policies in the original state-space are sufficient for attaining maximal reward, but need to be non-stationary. As for stationary Markov policies, we show they are sub-optimal in general. Consequently, we devise a non-stationary Q-learning style model-based algorithm that solves delayed execution tasks without resorting to state-augmentation. Experiments on tabular, physical, and Atari domains reveal that it converges quickly to high performance even for substantial delays, while standard approaches that either ignore the delay or rely on state-augmentation struggle or fail due to divergence. The code is available at github.com/galdl/rl_delay_basic and github.com/galdl/rl_delay_atari.

Guy Lutsker, Gal Sapir, Jordi Merino et al.

Understanding how human health changes over time, and why responses to interventions vary between individuals, remains a central challenge in medicine. Here we present HealthFormer, a decoder-only transformer that models the human physiological trajectory generatively, by training on data from the Human Phenotype Project, a multi-visit cohort of over 15,000 deeply phenotyped individuals. We tokenise each participant's health trajectory across 667 measurements spanning seven domains: blood biomarkers, body composition, sleep physiology, continuous glucose monitoring, gut microbiome, wearable-derived physiology, and behaviour and medication exposure. We train HealthFormer to forecast individual physiological trajectories across these domains, and from this single generative objective a range of clinically relevant tasks can be expressed as queries on the model. We show that, without task-specific training, HealthFormer transfers to four independent cohorts and improves prediction for 27 of 30 incident-disease and mortality endpoints, exceeding established clinical risk scores in every comparison. We further show that the model can simulate interventions in silico: in a held-out personalised-nutrition trial, intervention-conditioned predictions recover individual six-month biomarker changes (e.g., Pearson r = 0.78 for diastolic blood pressure). Across 41 randomised intervention-outcome comparisons drawn from published trials, our results show that the predicted direction of effect agrees in every case, and the predicted mean falls within the reported 95% confidence interval in 30 cases. We position HealthFormer as an initial health world model, from which forecasting, risk stratification, and intervention-conditioned simulation arise as queries, providing a basis for clinical digital twins.

Yihan Du, Anna Winnicki, Gal Dalal et al.

Reinforcement Learning from Human Feedback (RLHF) has achieved impressive empirical successes while relying on a small amount of human feedback. However, there is limited theoretical justification for this phenomenon. Additionally, most recent studies focus on value-based algorithms despite the recent empirical successes of policy-based algorithms. In this work, we consider an RLHF algorithm based on policy optimization (PO-RLHF). The algorithm is based on the popular Policy Cover-Policy Gradient (PC-PG) algorithm, which assumes knowledge of the reward function. In PO-RLHF, knowledge of the reward function is not assumed, and the algorithm uses trajectory-based comparison feedback to infer the reward function. We provide performance bounds for PO-RLHF with low query complexity, which provides insight into why a small amount of human feedback may be sufficient to achieve good performance with RLHF. A key novelty is a trajectory-level elliptical potential analysis, which bounds the reward estimation error when comparison feedback (rather than numerical reward observation) is given. We provide and analyze algorithms PG-RLHF and NN-PG-RLHF for two settings: linear and neural function approximation, respectively.

Navdeep Kumar, Yashaswini Murthy, Itai Shufaro et al.

We present the first finite time global convergence analysis of policy gradient in the context of infinite horizon average reward Markov decision processes (MDPs). Specifically, we focus on ergodic tabular MDPs with finite state and action spaces. Our analysis shows that the policy gradient iterates converge to the optimal policy at a sublinear rate of $O\left({\frac{1}{T}}\right),$ which translates to $O\left({\log(T)}\right)$ regret, where $T$ represents the number of iterations. Prior work on performance bounds for discounted reward MDPs cannot be extended to average reward MDPs because the bounds grow proportional to the fifth power of the effective horizon. Thus, our primary contribution is in proving that the policy gradient algorithm converges for average-reward MDPs and in obtaining finite-time performance guarantees. In contrast to the existing discounted reward performance bounds, our performance bounds have an explicit dependence on constants that capture the complexity of the underlying MDP. Motivated by this observation, we reexamine and improve the existing performance bounds for discounted reward MDPs. We also present simulations to empirically evaluate the performance of average reward policy gradient algorithm.

Ryan Park, Darren J. Hsu, C. Brian Roland et al.

Inverse folding models play an important role in structure-based design by predicting amino acid sequences that fold into desired reference structures. Models like ProteinMPNN, a message-passing encoder-decoder model, are trained to reliably produce new sequences from a reference structure. However, when applied to peptides, these models are prone to generating repetitive sequences that do not fold into the reference structure. To address this, we fine-tune ProteinMPNN to produce diverse and structurally consistent peptide sequences via Direct Preference Optimization (DPO). We derive two enhancements to DPO: online diversity regularization and domain-specific priors. Additionally, we develop a new understanding on improving diversity in decoder models. When conditioned on OpenFold generated structures, our fine-tuned models achieve state-of-the-art structural similarity scores, improving base ProteinMPNN by at least 8%. Compared to standard DPO, our regularized method achieves up to 20% higher sequence diversity with no loss in structural similarity score.

Yonatan Ashlag, Uri Koren, Mirco Mutti et al.

State entropy regularization has empirically shown better exploration and sample complexity in reinforcement learning (RL). However, its theoretical guarantees have not been studied. In this paper, we show that state entropy regularization improves robustness to structured and spatially correlated perturbations. These types of variation are common in transfer learning but often overlooked by standard robust RL methods, which typically focus on small, uncorrelated changes. We provide a comprehensive characterization of these robustness properties, including formal guarantees under reward and transition uncertainty, as well as settings where the method performs poorly. Much of our analysis contrasts state entropy with the widely used policy entropy regularization, highlighting their different benefits. Finally, from a practical standpoint, we illustrate that compared with policy entropy, the robustness advantages of state entropy are more sensitive to the number of rollouts used for policy evaluation.

Ofir Nabati, Bo Dai, Shie Mannor et al.

The effect of representation has been demonstrated in reinforcement learning, from both theoretical and empirical successes. However, the existing representation learning mainly induced from model learning aspects, misaligning with our RL tasks. This work introduces Spectral Bellman Representation, a novel framework derived from the Inherent Bellman Error (IBE) condition, which aligns with the fundamental structure of Bellman updates across a space of possible value functions, therefore, directly towards value-based RL. Our key insight is the discovery of a fundamental spectral relationship: under the zero-IBE condition, the transformation of a distribution of value functions by the Bellman operator is intrinsically linked to the feature covariance structure. This spectral connection yields a new, theoretically-grounded objective for learning state-action features that inherently capture this Bellman-aligned covariance. Our method requires a simple modification to existing algorithms. We demonstrate that our learned representations enable structured exploration, by aligning feature covariance with Bellman dynamics, and improve overall performance, particularly in challenging hard-exploration and long-horizon credit assignment tasks. Our framework naturally extends to powerful multi-step Bellman operators, further broadening its impact. Spectral Bellman Representation offers a principled and effective path toward learning more powerful and structurally sound representations for value-based reinforcement learning.

Uri Koren, Navdeep Kumar, Uri Gadot et al.

Classical policy gradient (PG) methods in reinforcement learning frequently converge to suboptimal local optima, a challenge exacerbated in large or complex environments. This work investigates Policy Gradient with Tree Search (PGTS), an approach that integrates an $m$-step lookahead mechanism to enhance policy optimization. We provide theoretical analysis demonstrating that increasing the tree search depth $m$-monotonically reduces the set of undesirable stationary points and, consequently, improves the worst-case performance of any resulting stationary policy. Critically, our analysis accommodates practical scenarios where policy updates are restricted to states visited by the current policy, rather than requiring updates across the entire state space. Empirical evaluations on diverse MDP structures, including Ladder, Tightrope, and Gridworld environments, illustrate PGTS's ability to exhibit "farsightedness," navigate challenging reward landscapes, escape local traps where standard PG fails, and achieve superior solutions.

Ido Greenberg, Piotr Sielski, Hugo Linsenmaier et al.

Vehicle Routing Problems (VRP) are an extension of the Traveling Salesperson Problem and are a fundamental NP-hard challenge in combinatorial optimization. Solving VRP in real-time at large scale has become critical in numerous applications, from growing markets like last-mile delivery to emerging use-cases like interactive logistics planning. In many applications, one has to repeatedly solve VRP instances drawn from the same distribution, yet current state-of-the-art solvers treat each instance on its own without leveraging previous examples. We introduce an optimization framework where a reinforcement learning agent is trained on prior instances and quickly generates initial solutions, which are then further optimized by a genetic algorithm. This framework, Evolutionary Algorithm with Reinforcement Learning Initialization (EARLI), consistently outperforms current state-of-the-art solvers across various time budgets. For example, EARLI handles vehicle routing with 500 locations within one second, 10x faster than current solvers for the same solution quality, enabling real-time and interactive routing at scale. EARLI can generalize to new data, as we demonstrate on real e-commerce delivery data of a previously unseen city. By combining reinforcement learning and genetic algorithms, our hybrid framework takes a step forward to closer interdisciplinary collaboration between AI and optimization communities towards real-time optimization in diverse domains.

Ron Vainshtein, Zohar Rimon, Shie Mannor et al.

Recent advancements in imitation learning have led to transformer-based behavior foundation models (BFMs) that enable multi-modal, human-like control for humanoid agents. While excelling at zero-shot generation of robust behaviors, BFMs often require meticulous prompt engineering for specific tasks, potentially yielding suboptimal results. We introduce "Task Tokens", a method to effectively tailor BFMs to specific tasks while preserving their flexibility. Our approach leverages the transformer architecture of BFMs to learn a new task-specific encoder through reinforcement learning, keeping the original BFM frozen. This allows incorporation of user-defined priors, balancing reward design and prompt engineering. By training a task encoder to map observations to tokens, used as additional BFM inputs, we guide performance improvement while maintaining the model's diverse control characteristics. We demonstrate Task Tokens' efficacy across various tasks, including out-of-distribution scenarios, and show their compatibility with other prompting modalities. Our results suggest that Task Tokens offer a promising approach for adapting BFMs to specific control tasks while retaining their generalization capabilities.

Navdeep Kumar, Adarsh Gupta, Maxence Mohamed Elfatihi et al.

We study robust Markov decision processes (RMDPs) with non-rectangular uncertainty sets, which capture interdependencies across states unlike traditional rectangular models. While non-rectangular robust policy evaluation is generally NP-hard, even in approximation, we identify a powerful class of $L_p$-bounded uncertainty sets that avoid these complexity barriers due to their structural simplicity. We further show that this class can be decomposed into infinitely many \texttt{sa}-rectangular $L_p$-bounded sets and leverage its structural properties to derive a novel dual formulation for $L_p$ RMDPs. This formulation provides key insights into the adversary's strategy and enables the development of the first robust policy evaluation algorithms for non-rectangular RMDPs. Empirical results demonstrate that our approach significantly outperforms brute-force methods, establishing a promising foundation for future investigation into non-rectangular robust MDPs.

Uri Gadot, Assaf Shocher, Shie Mannor et al.

Video encoders optimize compression for human perception by minimizing reconstruction error under bit-rate constraints. In many modern applications such as autonomous driving, an overwhelming majority of videos serve as input for AI systems performing tasks like object recognition or segmentation, rather than being watched by humans. It is therefore useful to optimize the encoder for a downstream task instead of for perceptual image quality. However, a major challenge is how to combine such downstream optimization with existing standard video encoders, which are highly efficient and popular. Here, we address this challenge by controlling the Quantization Parameters (QPs) at the macro-block level to optimize the downstream task. This granular control allows us to prioritize encoding for task-relevant regions within each frame. We formulate this optimization problem as a Reinforcement Learning (RL) task, where the agent learns to balance long-term implications of choosing QPs on both task performance and bit-rate constraints. Notably, our policy does not require the downstream task as an input during inference, making it suitable for streaming applications and edge devices such as vehicles. We demonstrate significant improvements in two tasks, car detection, and ROI (saliency) encoding. Our approach improves task performance for a given bit rate compared to traditional task agnostic encoding methods, paving the way for more efficient task-aware video compression.

Navdeep Kumar, Priyank Agrawal, Giorgia Ramponi et al.

We analyze the global convergence of the single-timescale actor-critic (AC) algorithm for the infinite-horizon discounted Markov Decision Processes (MDPs) with finite state spaces. To this end, we introduce an elegant analytical framework for handling complex, coupled recursions inherent in the algorithm. Leveraging this framework, we establish that the algorithm converges to an $ε$-close \textbf{globally optimal} policy with a sample complexity of \( O(ε^{-3}) \). This significantly improves upon the existing complexity of $O(ε^{-2})$ to achieve $ε$-close \textbf{stationary policy}, which is equivalent to the complexity of $O(ε^{-4})$ to achieve $ε$-close \textbf{globally optimal} policy using gradient domination lemma. Furthermore, we demonstrate that to achieve this improvement, the step sizes for both the actor and critic must decay as \( O(k^{-\frac{2}{3}}) \) with iteration $k$, diverging from the conventional \( O(k^{-\frac{1}{2}}) \) rates commonly used in (non)convex optimization.