Daniel J. Mankowitz

Jongmin Lee, Cosmin Paduraru, Daniel J. Mankowitz et al. · deepmind

We consider the offline constrained reinforcement learning (RL) problem, in which the agent aims to compute a policy that maximizes expected return while satisfying given cost constraints, learning only from a pre-collected dataset. This problem setting is appealing in many real-world scenarios, where direct interaction with the environment is costly or risky, and where the resulting policy should comply with safety constraints. However, it is challenging to compute a policy that guarantees satisfying the cost constraints in the offline RL setting, since the off-policy evaluation inherently has an estimation error. In this paper, we present an offline constrained RL algorithm that optimizes the policy in the space of the stationary distribution. Our algorithm, COptiDICE, directly estimates the stationary distribution corrections of the optimal policy with respect to returns, while constraining the cost upper bound, with the goal of yielding a cost-conservative policy for actual constraint satisfaction. Experimental results show that COptiDICE attains better policies in terms of constraint satisfaction and return-maximization, outperforming baseline algorithms.

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.

Jerry Luo, Cosmin Paduraru, Octavian Voicu et al. · deepmind

This paper is a technical overview of DeepMind and Google's recent work on reinforcement learning for controlling commercial cooling systems. Building on expertise that began with cooling Google's data centers more efficiently, we recently conducted live experiments on two real-world facilities in partnership with Trane Technologies, a building management system provider. These live experiments had a variety of challenges in areas such as evaluation, learning from offline data, and constraint satisfaction. Our paper describes these challenges in the hope that awareness of them will benefit future applied RL work. We also describe the way we adapted our RL system to deal with these challenges, resulting in energy savings of approximately 9% and 13% respectively at the two live experiment sites.

Amol Mandhane, Anton Zhernov, Maribeth Rauh et al.

Video streaming usage has seen a significant rise as entertainment, education, and business increasingly rely on online video. Optimizing video compression has the potential to increase access and quality of content to users, and reduce energy use and costs overall. In this paper, we present an application of the MuZero algorithm to the challenge of video compression. Specifically, we target the problem of learning a rate control policy to select the quantization parameters (QP) in the encoding process of libvpx, an open source VP9 video compression library widely used by popular video-on-demand (VOD) services. We treat this as a sequential decision making problem to maximize the video quality with an episodic constraint imposed by the target bitrate. Notably, we introduce a novel self-competition based reward mechanism to solve constrained RL with variable constraint satisfaction difficulty, which is challenging for existing constrained RL methods. We demonstrate that the MuZero-based rate control achieves an average 6.28% reduction in size of the compressed videos for the same delivered video quality level (measured as PSNR BD-rate) compared to libvpx's two-pass VBR rate control policy, while having better constraint satisfaction behavior.

Gabriel Dulac-Arnold, Nir Levine, Daniel J. Mankowitz et al.

Reinforcement learning (RL) has proven its worth in a series of artificial domains, and is beginning to show some successes in real-world scenarios. However, much of the research advances in RL are hard to leverage in real-world systems due to a series of assumptions that are rarely satisfied in practice. In this work, we identify and formalize a series of independent challenges that embody the difficulties that must be addressed for RL to be commonly deployed in real-world systems. For each challenge, we define it formally in the context of a Markov Decision Process, analyze the effects of the challenge on state-of-the-art learning algorithms, and present some existing attempts at tackling it. We believe that an approach that addresses our set of proposed challenges would be readily deployable in a large number of real world problems. Our proposed challenges are implemented in a suite of continuous control environments called the realworldrl-suite which we propose an as an open-source benchmark.

Gemini Team, Rohan Anil, Sebastian Borgeaud et al.

This report introduces a new family of multimodal models, Gemini, that exhibit remarkable capabilities across image, audio, video, and text understanding. The Gemini family consists of Ultra, Pro, and Nano sizes, suitable for applications ranging from complex reasoning tasks to on-device memory-constrained use-cases. Evaluation on a broad range of benchmarks shows that our most-capable Gemini Ultra model advances the state of the art in 30 of 32 of these benchmarks - notably being the first model to achieve human-expert performance on the well-studied exam benchmark MMLU, and improving the state of the art in every one of the 20 multimodal benchmarks we examined. We believe that the new capabilities of the Gemini family in cross-modal reasoning and language understanding will enable a wide variety of use cases. We discuss our approach toward post-training and deploying Gemini models responsibly to users through services including Gemini, Gemini Advanced, Google AI Studio, and Cloud Vertex AI.

Pengming Wang, Mikita Sazanovich, Berkin Ilbeyi et al.

Resource scheduling and allocation is a critical component of many high impact systems ranging from congestion control to cloud computing. Finding more optimal solutions to these problems often has significant impact on resource and time savings, reducing device wear-and-tear, and even potentially improving carbon emissions. In this paper, we focus on a specific instance of a scheduling problem, namely the memory mapping problem that occurs during compilation of machine learning programs: That is, mapping tensors to different memory layers to optimize execution time. We introduce an approach for solving the memory mapping problem using Reinforcement Learning. RL is a solution paradigm well-suited for sequential decision making problems that are amenable to planning, and combinatorial search spaces with high-dimensional data inputs. We formulate the problem as a single-player game, which we call the mallocGame, such that high-reward trajectories of the game correspond to efficient memory mappings on the target hardware. We also introduce a Reinforcement Learning agent, mallocMuZero, and show that it is capable of playing this game to discover new and improved memory mapping solutions that lead to faster execution times on real ML workloads on ML accelerators. We compare the performance of mallocMuZero to the default solver used by the Accelerated Linear Algebra (XLA) compiler on a benchmark of realistic ML workloads. In addition, we show that mallocMuZero is capable of improving the execution time of the recently published AlphaTensor matrix multiplication model.

Yujia Li, David Choi, Junyoung Chung et al.

Programming is a powerful and ubiquitous problem-solving tool. Developing systems that can assist programmers or even generate programs independently could make programming more productive and accessible, yet so far incorporating innovations in AI has proven challenging. Recent large-scale language models have demonstrated an impressive ability to generate code, and are now able to complete simple programming tasks. However, these models still perform poorly when evaluated on more complex, unseen problems that require problem-solving skills beyond simply translating instructions into code. For example, competitive programming problems which require an understanding of algorithms and complex natural language remain extremely challenging. To address this gap, we introduce AlphaCode, a system for code generation that can create novel solutions to these problems that require deeper reasoning. In simulated evaluations on recent programming competitions on the Codeforces platform, AlphaCode achieved on average a ranking of top 54.3% in competitions with more than 5,000 participants. We found that three key components were critical to achieve good and reliable performance: (1) an extensive and clean competitive programming dataset for training and evaluation, (2) large and efficient-to-sample transformer-based architectures, and (3) large-scale model sampling to explore the search space, followed by filtering based on program behavior to a small set of submissions.

Daniel J. Mankowitz, Dan A. Calian, Rae Jeong et al.

Many real-world physical control systems are required to satisfy constraints upon deployment. Furthermore, real-world systems are often subject to effects such as non-stationarity, wear-and-tear, uncalibrated sensors and so on. Such effects effectively perturb the system dynamics and can cause a policy trained successfully in one domain to perform poorly when deployed to a perturbed version of the same domain. This can affect a policy's ability to maximize future rewards as well as the extent to which it satisfies constraints. We refer to this as constrained model misspecification. We present an algorithm that mitigates this form of misspecification, and showcase its performance in multiple simulated Mujoco tasks from the Real World Reinforcement Learning (RWRL) suite.

Dan A. Calian, Daniel J. Mankowitz, Tom Zahavy et al.

Deploying Reinforcement Learning (RL) agents to solve real-world applications often requires satisfying complex system constraints. Often the constraint thresholds are incorrectly set due to the complex nature of a system or the inability to verify the thresholds offline (e.g, no simulator or reasonable offline evaluation procedure exists). This results in solutions where a task cannot be solved without violating the constraints. However, in many real-world cases, constraint violations are undesirable yet they are not catastrophic, motivating the need for soft-constrained RL approaches. We present a soft-constrained RL approach that utilizes meta-gradients to find a good trade-off between expected return and minimizing constraint violations. We demonstrate the effectiveness of this approach by showing that it consistently outperforms the baselines across four different MuJoCo domains.

Daniel J. Mankowitz, Nir Levine, Rae Jeong et al.

We provide a framework for incorporating robustness -- to perturbations in the transition dynamics which we refer to as model misspecification -- into continuous control Reinforcement Learning (RL) algorithms. We specifically focus on incorporating robustness into a state-of-the-art continuous control RL algorithm called Maximum a-posteriori Policy Optimization (MPO). We achieve this by learning a policy that optimizes for a worst case expected return objective and derive a corresponding robust entropy-regularized Bellman contraction operator. In addition, we introduce a less conservative, soft-robust, entropy-regularized objective with a corresponding Bellman operator. We show that both, robust and soft-robust policies, outperform their non-robust counterparts in nine Mujoco domains with environment perturbations. In addition, we show improved robust performance on a high-dimensional, simulated, dexterous robotic hand. Finally, we present multiple investigative experiments that provide a deeper insight into the robustness framework. This includes an adaptation to another continuous control RL algorithm as well as learning the uncertainty set from offline data. Performance videos can be found online at https://sites.google.com/view/robust-rl.

Chen Tessler, Tom Zahavy, Deborah Cohen et al.

We propose a computationally efficient algorithm that combines compressed sensing with imitation learning to solve text-based games with combinatorial action spaces. Specifically, we introduce a new compressed sensing algorithm, named IK-OMP, which can be seen as an extension to the Orthogonal Matching Pursuit (OMP). We incorporate IK-OMP into a supervised imitation learning setting and show that the combined approach (Sparse Imitation Learning, Sparse-IL) solves the entire text-based game of Zork1 with an action space of approximately 10 million actions given both perfect and noisy demonstrations.

Tom Zahavy, Matan Haroush, Nadav Merlis et al.

Learning how to act when there are many available actions in each state is a challenging task for Reinforcement Learning (RL) agents, especially when many of the actions are redundant or irrelevant. In such cases, it is sometimes easier to learn which actions not to take. In this work, we propose the Action-Elimination Deep Q-Network (AE-DQN) architecture that combines a Deep RL algorithm with an Action Elimination Network (AEN) that eliminates sub-optimal actions. The AEN is trained to predict invalid actions, supervised by an external elimination signal provided by the environment. Simulations demonstrate a considerable speedup and added robustness over vanilla DQN in text-based games with over a thousand discrete actions.

Chen Tessler, Daniel J. Mankowitz, Shie Mannor

Solving tasks in Reinforcement Learning is no easy feat. As the goal of the agent is to maximize the accumulated reward, it often learns to exploit loopholes and misspecifications in the reward signal resulting in unwanted behavior. While constraints may solve this issue, there is no closed form solution for general constraints. In this work we present a novel multi-timescale approach for constrained policy optimization, called `Reward Constrained Policy Optimization' (RCPO), which uses an alternative penalty signal to guide the policy towards a constraint satisfying one. We prove the convergence of our approach and provide empirical evidence of its ability to train constraint satisfying policies.

Esther Derman, Daniel J. Mankowitz, Timothy A. Mann et al.

Robust Reinforcement Learning aims to derive optimal behavior that accounts for model uncertainty in dynamical systems. However, previous studies have shown that by considering the worst case scenario, robust policies can be overly conservative. Our soft-robust framework is an attempt to overcome this issue. In this paper, we present a novel Soft-Robust Actor-Critic algorithm (SR-AC). It learns an optimal policy with respect to a distribution over an uncertainty set and stays robust to model uncertainty but avoids the conservativeness of robust strategies. We show the convergence of SR-AC and test the efficiency of our approach on different domains by comparing it against regular learning methods and their robust formulations.

Daniel J. Mankowitz, Augustin Žídek, André Barreto et al.

Some real-world domains are best characterized as a single task, but for others this perspective is limiting. Instead, some tasks continually grow in complexity, in tandem with the agent's competence. In continual learning, also referred to as lifelong learning, there are no explicit task boundaries or curricula. As learning agents have become more powerful, continual learning remains one of the frontiers that has resisted quick progress. To test continual learning capabilities we consider a challenging 3D domain with an implicit sequence of tasks and sparse rewards. We propose a novel agent architecture called Unicorn, which demonstrates strong continual learning and outperforms several baseline agents on the proposed domain. The agent achieves this by jointly representing and learning multiple policies efficiently, using a parallel off-policy learning setup.

Daniel J. Mankowitz, Timothy A. Mann, Pierre-Luc Bacon et al.

Robust reinforcement learning aims to produce policies that have strong guarantees even in the face of environments/transition models whose parameters have strong uncertainty. Existing work uses value-based methods and the usual primitive action setting. In this paper, we propose robust methods for learning temporally abstract actions, in the framework of options. We present a Robust Options Policy Iteration (ROPI) algorithm with convergence guarantees, which learns options that are robust to model uncertainty. We utilize ROPI to learn robust options with the Robust Options Deep Q Network (RO-DQN) that solves multiple tasks and mitigates model misspecification due to model uncertainty. We present experimental results which suggest that policy iteration with linear features may have an inherent form of robustness when using coarse feature representations. In addition, we present experimental results which demonstrate that robustness helps policy iteration implemented on top of deep neural networks to generalize over a much broader range of dynamics than non-robust policy iteration.

Daniel J. Mankowitz, Aviv Tamar, Shie Mannor

Hierarchical abstractions, also known as options -- a type of temporally extended action (Sutton et. al. 1999) that enables a reinforcement learning agent to plan at a higher level, abstracting away from the lower-level details. In this work, we learn reusable options whose parameters can vary, encouraging different behaviors, based on the current situation. In principle, these behaviors can include vigor, defence or even risk-averseness. These are some examples of what we refer to in the broader context as Situational Awareness (SA). We incorporate SA, in the form of vigor, into hierarchical RL by defining and learning situationally aware options in a Probabilistic Goal Semi-Markov Decision Process (PG-SMDP). This is achieved using our Situationally Aware oPtions (SAP) policy gradient algorithm which comes with a theoretical convergence guarantee. We learn reusable options in different scenarios in a RoboCup soccer domain (i.e., winning/losing). These options learn to execute with different levels of vigor resulting in human-like behaviours such as `time-wasting' in the winning scenario. We show the potential of the agent to exit bad local optima using reusable options in RoboCup. Finally, using SAP, the agent mitigates feature-based model misspecification in a Bottomless Pit of Death domain.

Nir Levine, Tom Zahavy, Daniel J. Mankowitz et al.

Deep reinforcement learning (DRL) methods such as the Deep Q-Network (DQN) have achieved state-of-the-art results in a variety of challenging, high-dimensional domains. This success is mainly attributed to the power of deep neural networks to learn rich domain representations for approximating the value function or policy. Batch reinforcement learning methods with linear representations, on the other hand, are more stable and require less hyper parameter tuning. Yet, substantial feature engineering is necessary to achieve good results. In this work we propose a hybrid approach -- the Least Squares Deep Q-Network (LS-DQN), which combines rich feature representations learned by a DRL algorithm with the stability of a linear least squares method. We do this by periodically re-training the last hidden layer of a DRL network with a batch least squares update. Key to our approach is a Bayesian regularization term for the least squares update, which prevents over-fitting to the more recent data. We tested LS-DQN on five Atari games and demonstrate significant improvement over vanilla DQN and Double-DQN. We also investigated the reasons for the superior performance of our method. Interestingly, we found that the performance improvement can be attributed to the large batch size used by the LS method when optimizing the last layer.

Daniel J. Mankowitz, Aviv Tamar, Shie Mannor

Hierarchical Reinforcement Learning has been previously shown to speed up the convergence rate of RL planning algorithms as well as mitigate feature-based model misspecification (Mankowitz et. al. 2016a,b, Bacon 2015). To do so, it utilizes hierarchical abstractions, also known as skills -- a type of temporally extended action (Sutton et. al. 1999) to plan at a higher level, abstracting away from the lower-level details. We incorporate risk sensitivity, also referred to as Situational Awareness (SA), into hierarchical RL for the first time by defining and learning risk aware skills in a Probabilistic Goal Semi-Markov Decision Process (PG-SMDP). This is achieved using our novel Situational Awareness by Risk-Conscious Skills (SARiCoS) algorithm which comes with a theoretical convergence guarantee. We show in a RoboCup soccer domain that the learned risk aware skills exhibit complex human behaviors such as `time-wasting' in a soccer game. In addition, the learned risk aware skills are able to mitigate reward-based model misspecification.

Chen Tessler, Shahar Givony, Tom Zahavy et al.

We propose a lifelong learning system that has the ability to reuse and transfer knowledge from one task to another while efficiently retaining the previously learned knowledge-base. Knowledge is transferred by learning reusable skills to solve tasks in Minecraft, a popular video game which is an unsolved and high-dimensional lifelong learning problem. These reusable skills, which we refer to as Deep Skill Networks, are then incorporated into our novel Hierarchical Deep Reinforcement Learning Network (H-DRLN) architecture using two techniques: (1) a deep skill array and (2) skill distillation, our novel variation of policy distillation (Rusu et. al. 2015) for learning skills. Skill distillation enables the HDRLN to efficiently retain knowledge and therefore scale in lifelong learning, by accumulating knowledge and encapsulating multiple reusable skills into a single distilled network. The H-DRLN exhibits superior performance and lower learning sample complexity compared to the regular Deep Q Network (Mnih et. al. 2015) in sub-domains of Minecraft.

Daniel J. Mankowitz, Timothy A. Mann, Shie Mannor

We introduce the Adaptive Skills, Adaptive Partitions (ASAP) framework that (1) learns skills (i.e., temporally extended actions or options) as well as (2) where to apply them. We believe that both (1) and (2) are necessary for a truly general skill learning framework, which is a key building block needed to scale up to lifelong learning agents. The ASAP framework can also solve related new tasks simply by adapting where it applies its existing learned skills. We prove that ASAP converges to a local optimum under natural conditions. Finally, our experimental results, which include a RoboCup domain, demonstrate the ability of ASAP to learn where to reuse skills as well as solve multiple tasks with considerably less experience than solving each task from scratch.

Daniel J. Mankowitz, Timothy A. Mann, Shie Mannor

For complex, high-dimensional Markov Decision Processes (MDPs), it may be necessary to represent the policy with function approximation. A problem is misspecified whenever, the representation cannot express any policy with acceptable performance. We introduce IHOMP : an approach for solving misspecified problems. IHOMP iteratively learns a set of context specialized options and combines these options to solve an otherwise misspecified problem. Our main contribution is proving that IHOMP enjoys theoretical convergence guarantees. In addition, we extend IHOMP to exploit Option Interruption (OI) enabling it to decide where the learned options can be reused. Our experiments demonstrate that IHOMP can find near-optimal solutions to otherwise misspecified problems and that OI can further improve the solutions.

Daniel J. Mankowitz, Ehud Rivlin

We present a novel Visual Odometry algorithm entitled Circular FREAK-ORB (CFORB). This algorithm detects features using the well-known ORB algorithm [12] and computes feature descriptors using the FREAK algorithm [14]. CFORB is invariant to both rotation and scale changes, and is suitable for use in environments with uneven terrain. Two visual geometric constraints have been utilized in order to remove invalid feature descriptor matches. These constraints have not previously been utilized in a Visual Odometry algorithm. A variation to circular matching [16] has also been implemented. This allows features to be matched between images without having to be dependent upon the epipolar constraint. This algorithm has been run on the KITTI benchmark dataset and achieves a competitive average translational error of $3.73 \%$ and average rotational error of $0.0107 deg/m$. CFORB has also been run in an indoor environment and achieved an average translational error of $3.70 \%$. After running CFORB in a highly textured environment with an approximately uniform feature spread across the images, the algorithm achieves an average translational error of $2.4 \%$ and an average rotational error of $0.009 deg/m$.

Daniel J. Mankowitz, Timothy A. Mann, Shie Mannor

The monolithic approach to policy representation in Markov Decision Processes (MDPs) looks for a single policy that can be represented as a function from states to actions. For the monolithic approach to succeed (and this is not always possible), a complex feature representation is often necessary since the policy is a complex object that has to prescribe what actions to take all over the state space. This is especially true in large domains with complicated dynamics. It is also computationally inefficient to both learn and plan in MDPs using a complex monolithic approach. We present a different approach where we restrict the policy space to policies that can be represented as combinations of simpler, parameterized skills---a type of temporally extended action, with a simple policy representation. We introduce Learning Skills via Bootstrapping (LSB) that can use a broad family of Reinforcement Learning (RL) algorithms as a "black box" to iteratively learn parametrized skills. Initially, the learned skills are short-sighted but each iteration of the algorithm allows the skills to bootstrap off one another, improving each skill in the process. We prove that this bootstrapping process returns a near-optimal policy. Furthermore, our experiments demonstrate that LSB can solve MDPs that, given the same representational power, could not be solved by a monolithic approach. Thus, planning with learned skills results in better policies without requiring complex policy representations.