Nevena Lazic

Brendan D. Tracey, Andrea Michi, Yuri Chervonyi et al. · deepmind

Reinforcement learning (RL) has shown promising results for real-time control systems, including the domain of plasma magnetic control. However, there are still significant drawbacks compared to traditional feedback control approaches for magnetic confinement. In this work, we address key drawbacks of the RL method; achieving higher control accuracy for desired plasma properties, reducing the steady-state error, and decreasing the required time to learn new tasks. We build on top of \cite{degrave2022magnetic}, and present algorithmic improvements to the agent architecture and training procedure. We present simulation results that show up to 65\% improvement in shape accuracy, achieve substantial reduction in the long-term bias of the plasma current, and additionally reduce the training time required to learn new tasks by a factor of 3 or more. We present new experiments using the upgraded RL-based controllers on the TCV tokamak, which validate the simulation results achieved, and point the way towards routinely achieving accurate discharges using the RL approach.

Dong Yin, Sridhar Thiagarajan, Nevena Lazic et al. · deepmind

The focus of this work is sample-efficient deep reinforcement learning (RL) with a simulator. One useful property of simulators is that it is typically easy to reset the environment to a previously observed state. We propose an algorithmic framework, named uncertainty-first local planning (UFLP), that takes advantage of this property. Concretely, in each data collection iteration, with some probability, our meta-algorithm resets the environment to an observed state which has high uncertainty, instead of sampling according to the initial-state distribution. The agent-environment interaction then proceeds as in the standard online RL setting. We demonstrate that this simple procedure can dramatically improve the sample cost of several baseline RL algorithms on difficult exploration tasks. Notably, with our framework, we can achieve super-human performance on the notoriously hard Atari game, Montezuma's Revenge, with a simple (distributional) double DQN. Our work can be seen as an efficient approximate implementation of an existing algorithm with theoretical guarantees, which offers an interpretation of the positive empirical results.

David B. D'Ambrosio, Jonathan Abelian, Saminda Abeyruwan et al.

We present a deep-dive into a real-world robotic learning system that, in previous work, was shown to be capable of hundreds of table tennis rallies with a human and has the ability to precisely return the ball to desired targets. This system puts together a highly optimized perception subsystem, a high-speed low-latency robot controller, a simulation paradigm that can prevent damage in the real world and also train policies for zero-shot transfer, and automated real world environment resets that enable autonomous training and evaluation on physical robots. We complement a complete system description, including numerous design decisions that are typically not widely disseminated, with a collection of studies that clarify the importance of mitigating various sources of latency, accounting for training and deployment distribution shifts, robustness of the perception system, sensitivity to policy hyper-parameters, and choice of action space. A video demonstrating the components of the system and details of experimental results can be found at https://youtu.be/uFcnWjB42I0.

Hongzi Mao, Chenjie Gu, Miaosen Wang et al.

In modern video encoders, rate control is a critical component and has been heavily engineered. It decides how many bits to spend to encode each frame, in order to optimize the rate-distortion trade-off over all video frames. This is a challenging constrained planning problem because of the complex dependency among decisions for different video frames and the bitrate constraint defined at the end of the episode. We formulate the rate control problem as a Partially Observable Markov Decision Process (POMDP), and apply imitation learning to learn a neural rate control policy. We demonstrate that by learning from optimal video encoding trajectories obtained through evolution strategies, our learned policy achieves better encoding efficiency and has minimal constraint violation. In addition to imitating the optimal actions, we find that additional auxiliary losses, data augmentation/refinement and inference-time policy improvements are critical for learning a good rate control policy. We evaluate the learned policy against the rate control policy in libvpx, a widely adopted open source VP9 codec library, in the two-pass variable bitrate (VBR) mode. We show that over a diverse set of real-world videos, our learned policy achieves 8.5% median bitrate reduction without sacrificing video quality.

Alan Malek, Jiawei Ge, Nevena Lazic et al.

While state-of-the-art large language models (LLMs) demonstrate advanced reasoning capabilities-achieving remarkable performance on challenging competitive math and coding benchmarks-they also frequently fail on tasks that are easy for humans. This work studies the performance of frontier LLMs on a broad set of such "easy" reasoning problems. By extending previous work in the literature, we create a suite of procedurally generated simple reasoning tasks, including counting, first-order logic, proof trees, and travel planning, with changeable parameters (such as document length. or the number of variables in a math problem) that can arbitrarily increase the amount of computation required to produce the answer while preserving the fundamental difficulty. While previous work showed that traditional, non-thinking models can be made to fail on such problems, we demonstrate that even state-of-the-art thinking models consistently fail on such problems and for similar reasons (e.g. statistical shortcuts, errors in intermediate steps, and difficulties in processing long contexts). To further understand the behavior of the models, we introduce the unpuzzles dataset, a different "easy" benchmark consisting of trivialized versions of well-known math and logic puzzles. Interestingly, while modern LLMs excel at solving the original puzzles, they tend to fail on the trivialized versions, exhibiting several systematic failure patterns related to memorizing the originals. We show that this happens even if the models are otherwise able to solve problems with different descriptions but requiring the same logic. Our results highlight that out-of-distribution generalization is still problematic for frontier language models and the new generation of thinking models, even for simple reasoning tasks, and making tasks easier does not necessarily imply improved performance.

Yasin Abbasi-Yadkori, Andras Gyorgy, Nevena Lazic

We study the non-stationary stochastic multi-armed bandit problem, where the reward statistics of each arm may change several times during the course of learning. The performance of a learning algorithm is evaluated in terms of their dynamic regret, which is defined as the difference between the expected cumulative reward of an agent choosing the optimal arm in every time step and the cumulative reward of the learning algorithm. One way to measure the hardness of such environments is to consider how many times the identity of the optimal arm can change. We propose a method that achieves, in $K$-armed bandit problems, a near-optimal $\widetilde O(\sqrt{K N(S+1)})$ dynamic regret, where $N$ is the time horizon of the problem and $S$ is the number of times the identity of the optimal arm changes, without prior knowledge of $S$. Previous works for this problem obtain regret bounds that scale with the number of changes (or the amount of change) in the reward functions, which can be much larger, or assume prior knowledge of $S$ to achieve similar bounds.

Nevena Lazic, Dong Yin, Yasin Abbasi-Yadkori et al.

In this work, we study algorithms for learning in infinite-horizon undiscounted Markov decision processes (MDPs) with function approximation. We first show that the regret analysis of the Politex algorithm (a version of regularized policy iteration) can be sharpened from $O(T^{3/4})$ to $O(\sqrt{T})$ under nearly identical assumptions, and instantiate the bound with linear function approximation. Our result provides the first high-probability $O(\sqrt{T})$ regret bound for a computationally efficient algorithm in this setting. The exact implementation of Politex with neural network function approximation is inefficient in terms of memory and computation. Since our analysis suggests that we need to approximate the average of the action-value functions of past policies well, we propose a simple efficient implementation where we train a single Q-function on a replay buffer with past data. We show that this often leads to superior performance over other implementation choices, especially in terms of wall-clock time. Our work also provides a novel theoretical justification for using experience replay within policy iteration algorithms.

Nevena Lazic, Dong Yin, Mehrdad Farajtabar et al.

This work focuses on off-policy evaluation (OPE) with function approximation in infinite-horizon undiscounted Markov decision processes (MDPs). For MDPs that are ergodic and linear (i.e. where rewards and dynamics are linear in some known features), we provide the first finite-sample OPE error bound, extending existing results beyond the episodic and discounted cases. In a more general setting, when the feature dynamics are approximately linear and for arbitrary rewards, we propose a new approach for estimating stationary distributions with function approximation. We formulate this problem as finding the maximum-entropy distribution subject to matching feature expectations under empirical dynamics. We show that this results in an exponential-family distribution whose sufficient statistics are the features, paralleling maximum-entropy approaches in supervised learning. We demonstrate the effectiveness of the proposed OPE approaches in multiple environments.

Wenbo Gao, Laura Graesser, Krzysztof Choromanski et al.

We propose a model-free algorithm for learning efficient policies capable of returning table tennis balls by controlling robot joints at a rate of 100Hz. We demonstrate that evolutionary search (ES) methods acting on CNN-based policy architectures for non-visual inputs and convolving across time learn compact controllers leading to smooth motions. Furthermore, we show that with appropriately tuned curriculum learning on the task and rewards, policies are capable of developing multi-modal styles, specifically forehand and backhand stroke, whilst achieving 80\% return rate on a wide range of ball throws. We observe that multi-modality does not require any architectural priors, such as multi-head architectures or hierarchical policies.

Botao Hao, Nevena Lazic, Yasin Abbasi-Yadkori et al.

Model-free reinforcement learning algorithms combined with value function approximation have recently achieved impressive performance in a variety of application domains. However, the theoretical understanding of such algorithms is limited, and existing results are largely focused on episodic or discounted Markov decision processes (MDPs). In this work, we present adaptive approximate policy iteration (AAPI), a learning scheme which enjoys a $\tilde{O}(T^{2/3})$ regret bound for undiscounted, continuing learning in uniformly ergodic MDPs. This is an improvement over the best existing bound of $\tilde{O}(T^{3/4})$ for the average-reward case with function approximation. Our algorithm and analysis rely on online learning techniques, where value functions are treated as losses. The main technical novelty is the use of a data-dependent adaptive learning rate coupled with a so-called optimistic prediction of upcoming losses. In addition to theoretical guarantees, we demonstrate the advantages of our approach empirically on several environments.

Yasin Abbasi-Yadkori, Nevena Lazic, Csaba Szepesvari et al.

We study algorithms for average-cost reinforcement learning problems with value function approximation. Our starting point is the recently proposed POLITEX algorithm, a version of policy iteration where the policy produced in each iteration is near-optimal in hindsight for the sum of all past value function estimates. POLITEX has sublinear regret guarantees in uniformly-mixing MDPs when the value estimation error can be controlled, which can be satisfied if all policies sufficiently explore the environment. Unfortunately, this assumption is often unrealistic. Motivated by the rapid growth of interest in developing policies that learn to explore their environment in the lack of rewards (also known as no-reward learning), we replace the previous assumption that all policies explore the environment with that a single, sufficiently exploring policy is available beforehand. The main contribution of the paper is the modification of POLITEX to incorporate such an exploration policy in a way that allows us to obtain a regret guarantee similar to the previous one but without requiring that all policies explore environment. In addition to the novel theoretical guarantees, we demonstrate the benefits of our scheme on environments which are difficult to explore using simple schemes like dithering. While the solution we obtain may not achieve the best possible regret, it is the first result that shows how to control the regret in the presence of function approximation errors on problems where exploration is nontrivial. Our approach can also be seen as a way of reducing the problem of minimizing the regret to learning a good exploration policy. We believe that modular approaches like ours can be highly beneficial in tackling harder control problems.

Reza Mahjourian, Risto Miikkulainen, Nevena Lazic et al.

Training robots with physical bodies requires developing new methods and action representations that allow the learning agents to explore the space of policies efficiently. This work studies sample-efficient learning of complex policies in the context of robot table tennis. It incorporates learning into a hierarchical control framework using a model-free strategy layer (which requires complex reasoning about opponents that is difficult to do in a model-based way), model-based prediction of external objects (which are difficult to control directly with analytic control methods, but governed by learnable and relatively simple laws of physics), and analytic controllers for the robot itself. Human demonstrations are used to train dynamics models, which together with the analytic controller allow any robot that is physically capable to play table tennis without training episodes. Using only about 7,000 demonstrated trajectories, a striking policy can hit ball targets with about 20 cm error. Self-play is used to train cooperative and adversarial strategies on top of model-based striking skills trained from human demonstrations. After only about 24,000 strikes in self-play the agent learns to best exploit the human dynamics models for longer cooperative games. Further experiments demonstrate that more flexible variants of the policy can discover new strikes not demonstrated by humans and achieve higher performance at the expense of lower sample-efficiency. Experiments are carried out in a virtual reality environment using sensory observations that are obtainable in the real world. The high sample-efficiency demonstrated in the evaluations show that the proposed method is suitable for learning directly on physical robots without transfer of models or policies from simulation. Supplementary material available at https://sites.google.com/view/robottabletennis

Alon Cohen, Avinatan Hassidim, Tomer Koren et al.

We study the problem of controlling linear time-invariant systems with known noisy dynamics and adversarially chosen quadratic losses. We present the first efficient online learning algorithms in this setting that guarantee $O(\sqrt{T})$ regret under mild assumptions, where $T$ is the time horizon. Our algorithms rely on a novel SDP relaxation for the steady-state distribution of the system. Crucially, and in contrast to previously proposed relaxations, the feasible solutions of our SDP all correspond to "strongly stable" policies that mix exponentially fast to a steady state.

Yasin Abbasi-Yadkori, Nevena Lazic, Csaba Szepesvari

Model-free approaches for reinforcement learning (RL) and continuous control find policies based only on past states and rewards, without fitting a model of the system dynamics. They are appealing as they are general purpose and easy to implement; however, they also come with fewer theoretical guarantees than model-based RL. In this work, we present a new model-free algorithm for controlling linear quadratic (LQ) systems, and show that its regret scales as $O(T^{ξ+2/3})$ for any small $ξ>0$ if time horizon satisfies $T>C^{1/ξ}$ for a constant $C$. The algorithm is based on a reduction of control of Markov decision processes to an expert prediction problem. In practice, it corresponds to a variant of policy iteration with forced exploration, where the policy in each phase is greedy with respect to the average of all previous value functions. This is the first model-free algorithm for adaptive control of LQ systems that provably achieves sublinear regret and has a polynomial computation cost. Empirically, our algorithm dramatically outperforms standard policy iteration, but performs worse than a model-based approach.

Amit Daniely, Nevena Lazic, Yoram Singer et al.

High-dimensional sparse data present computational and statistical challenges for supervised learning. We propose compact linear sketches for reducing the dimensionality of the input, followed by a single layer neural network. We show that any sparse polynomial function can be computed, on nearly all sparse binary vectors, by a single layer neural network that takes a compact sketch of the vector as input. Consequently, when a set of sparse binary vectors is approximately separable using a sparse polynomial, there exists a single-layer neural network that takes a short sketch as input and correctly classifies nearly all the points. Previous work has proposed using sketches to reduce dimensionality while preserving the hypothesis class. However, the sketch size has an exponential dependence on the degree in the case of polynomial classifiers. In stark contrast, our approach of using improper learning, using a larger hypothesis class allows the sketch size to have a logarithmic dependence on the degree. Even in the linear case, our approach allows us to improve on the pesky $O({1}/{γ^2})$ dependence of random projections, on the margin $γ$. We empirically show that our approach leads to more compact neural networks than related methods such as feature hashing at equal or better performance.

Dan Gillick, Nevena Lazic, Kuzman Ganchev et al.

Entity type tagging is the task of assigning category labels to each mention of an entity in a document. While standard systems focus on a small set of types, recent work (Ling and Weld, 2012) suggests that using a large fine-grained label set can lead to dramatic improvements in downstream tasks. In the absence of labeled training data, existing fine-grained tagging systems obtain examples automatically, using resolved entities and their types extracted from a knowledge base. However, since the appropriate type often depends on context (e.g. Washington could be tagged either as city or government), this procedure can result in spurious labels, leading to poorer generalization. We propose the task of context-dependent fine type tagging, where the set of acceptable labels for a mention is restricted to only those deducible from the local context (e.g. sentence or document). We introduce new resources for this task: 12,017 mentions annotated with their context-dependent fine types, and we provide baseline experimental results on this data.