Lukas Schäfer

Lukas Schäfer, Filippos Christianos, Amos Storkey et al. · microsoft-research

Successful deployment of multi-agent reinforcement learning often requires agents to adapt their behaviour. In this work, we discuss the problem of teamwork adaptation in which a team of agents needs to adapt their policies to solve novel tasks with limited fine-tuning. Motivated by the intuition that agents need to be able to identify and distinguish tasks in order to adapt their behaviour to the current task, we propose to learn multi-agent task embeddings (MATE). These task embeddings are trained using an encoder-decoder architecture optimised for reconstruction of the transition and reward functions which uniquely identify tasks. We show that a team of agents is able to adapt to novel tasks when provided with task embeddings. We propose three MATE training paradigms: independent MATE, centralised MATE, and mixed MATE which vary in the information used for the task encoding. We show that the embeddings learned by MATE identify tasks and provide useful information which agents leverage during adaptation to novel tasks.

Alain Andres, Lukas Schäfer, Stefano V. Albrecht et al. · microsoft-research

One of the key challenges of Reinforcement Learning (RL) is the ability of agents to generalise their learned policy to unseen settings. Moreover, training RL agents requires large numbers of interactions with the environment. Motivated by the recent success of Offline RL and Imitation Learning (IL), we conduct a study to investigate whether agents can leverage offline data in the form of trajectories to improve the sample-efficiency in procedurally generated environments. We consider two settings of using IL from offline data for RL: (1) pre-training a policy before online RL training and (2) concurrently training a policy with online RL and IL from offline data. We analyse the impact of the quality (optimality of trajectories) and diversity (number of trajectories and covered level) of available offline trajectories on the effectiveness of both approaches. Across four well-known sparse reward tasks in the MiniGrid environment, we find that using IL for pre-training and concurrently during online RL training both consistently improve the sample-efficiency while converging to optimal policies. Furthermore, we show that pre-training a policy from as few as two trajectories can make the difference between learning an optimal policy at the end of online training and not learning at all. Our findings motivate the widespread adoption of IL for pre-training and concurrent IL in procedurally generated environments whenever offline trajectories are available or can be generated.

Aleksandar Krnjaic, Raul D. Steleac, Jonathan D. Thomas et al. · microsoft-research

We consider a warehouse in which dozens of mobile robots and human pickers work together to collect and deliver items within the warehouse. The fundamental problem we tackle, called the order-picking problem, is how these worker agents must coordinate their movement and actions in the warehouse to maximise performance in this task. Established industry methods using heuristic approaches require large engineering efforts to optimise for innately variable warehouse configurations. In contrast, multi-agent reinforcement learning (MARL) can be flexibly applied to diverse warehouse configurations (e.g. size, layout, number/types of workers, item replenishment frequency), and different types of order-picking paradigms (e.g. Goods-to-Person and Person-to-Goods), as the agents can learn how to cooperate optimally through experience. We develop hierarchical MARL algorithms in which a manager agent assigns goals to worker agents, and the policies of the manager and workers are co-trained toward maximising a global objective (e.g. pick rate). Our hierarchical algorithms achieve significant gains in sample efficiency over baseline MARL algorithms and overall pick rates over multiple established industry heuristics in a diverse set of warehouse configurations and different order-picking paradigms.

Ibrahim H. Ahmed, Cillian Brewitt, Ignacio Carlucho et al. · microsoft-research

The development of autonomous agents which can interact with other agents to accomplish a given task is a core area of research in artificial intelligence and machine learning. Towards this goal, the Autonomous Agents Research Group develops novel machine learning algorithms for autonomous systems control, with a specific focus on deep reinforcement learning and multi-agent reinforcement learning. Research problems include scalable learning of coordinated agent policies and inter-agent communication; reasoning about the behaviours, goals, and composition of other agents from limited observations; and sample-efficient learning based on intrinsic motivation, curriculum learning, causal inference, and representation learning. This article provides a broad overview of the ongoing research portfolio of the group and discusses open problems for future directions.

Trevor McInroe, Lukas Schäfer, Stefano V. Albrecht · microsoft-research

Learning control from pixels is difficult for reinforcement learning (RL) agents because representation learning and policy learning are intertwined. Previous approaches remedy this issue with auxiliary representation learning tasks, but they either do not consider the temporal aspect of the problem or only consider single-step transitions, which may cause learning inefficiencies if important environmental changes take many steps to manifest. We propose Hierarchical $k$-Step Latent (HKSL), an auxiliary task that learns multiple representations via a hierarchy of forward models that learn to communicate and an ensemble of $n$-step critics that all operate at varying magnitudes of step skipping. We evaluate HKSL in a suite of 30 robotic control tasks with and without distractors and a task of our creation. We find that HKSL either converges to higher or optimal episodic returns more quickly than several alternative representation learning approaches. Furthermore, we find that HKSL's representations capture task-relevant details accurately across timescales (even in the presence of distractors) and that communication channels between hierarchy levels organize information based on both sides of the communication process, both of which improve sample efficiency.

Lukas Schäfer, Oliver Slumbers, Stephen McAleer et al. · microsoft-research

Multi-agent reinforcement learning (MARL) requires agents to explore within a vast joint action space to find joint actions that lead to coordination. Existing value-based MARL algorithms commonly rely on random exploration, such as $ε$-greedy, to explore the environment which is not systematic and inefficient at identifying effective actions in multi-agent problems. Additionally, the concurrent training of the policies of multiple agents during training can render the optimisation non-stationary. This can lead to unstable value estimates, highly variant gradients, and ultimately hinder coordination between agents. To address these challenges, we propose ensemble value functions for multi-agent exploration (EMAX). EMAX is a framework to seamlessly extend value-based MARL algorithms. EMAX leverages an ensemble of value functions for each agent to guide their exploration, reduce the variance of their optimisation, and makes their policies more robust to miscoordination. EMAX achieves these benefits by (1) systematically guiding the exploration of agents with a UCB policy towards parts of the environment that require multiple agents to coordinate. (2) EMAX computes average value estimates across the ensemble as target values to reduce the variance of gradients and make optimisation more stable. (3) During evaluation, EMAX selects actions following a majority vote across the ensemble to reduce the likelihood of miscoordination. We first instantiate independent DQN with EMAX and evaluate it in 11 general-sum tasks with sparse rewards. We show that EMAX improves final evaluation returns by 185% across all tasks. We then evaluate EMAX on top of IDQN, VDN and QMIX in 21 common-reward tasks, and show that EMAX improves sample efficiency and final evaluation returns across all tasks over all three vanilla algorithms by 60%, 47%, and 538%, respectively.

Hanna Krasowski, Jakob Thumm, Marlon Müller et al.

Ensuring the safety of reinforcement learning (RL) algorithms is crucial to unlock their potential for many real-world tasks. However, vanilla RL and most safe RL approaches do not guarantee safety. In recent years, several methods have been proposed to provide hard safety guarantees for RL, which is essential for applications where unsafe actions could have disastrous consequences. Nevertheless, there is no comprehensive comparison of these provably safe RL methods. Therefore, we introduce a categorization of existing provably safe RL methods, present the conceptual foundations for both continuous and discrete action spaces, and empirically benchmark existing methods. We categorize the methods based on how they adapt the action: action replacement, action projection, and action masking. Our experiments on an inverted pendulum and a quadrotor stabilization task indicate that action replacement is the best-performing approach for these applications despite its comparatively simple realization. Furthermore, adding a reward penalty, every time the safety verification is engaged, improved training performance in our experiments. Finally, we provide practical guidance on selecting provably safe RL approaches depending on the safety specification, RL algorithm, and type of action space.

Lukas Schäfer, Pallavi Choudhury, Abdelhak Lemkhenter et al.

Behavior cloning (BC) is a practical offline imitation learning method, but it often fails when expert demonstrations are limited. Recent works have introduced a class of architectures named predictive inverse dynamics models (PIDM) that combine a future state predictor with an inverse dynamics model (IDM). While PIDM often outperforms BC, the reasons behind its benefits remain unclear. In this paper, we provide a theoretical explanation: PIDM introduces a bias-variance tradeoff. While predicting the future state introduces bias, conditioning the IDM on the prediction can significantly reduce variance. We establish conditions on the state predictor bias for PIDM to achieve lower prediction error and higher sample efficiency than BC, with the gap widening when additional data sources are available. We validate the theoretical insights empirically in 2D navigation tasks, where BC requires up to five times (three times on average) more demonstrations than PIDM to reach comparable performance; and in a complex 3D environment in a modern video game with high-dimensional visual inputs and stochastic transitions, where BC requires over 66\% more samples than PIDM.

Georgios Papoudakis, Filippos Christianos, Lukas Schäfer et al.

Multi-agent deep reinforcement learning (MARL) suffers from a lack of commonly-used evaluation tasks and criteria, making comparisons between approaches difficult. In this work, we provide a systematic evaluation and comparison of three different classes of MARL algorithms (independent learning, centralised multi-agent policy gradient, value decomposition) in a diverse range of cooperative multi-agent learning tasks. Our experiments serve as a reference for the expected performance of algorithms across different learning tasks, and we provide insights regarding the effectiveness of different learning approaches. We open-source EPyMARL, which extends the PyMARL codebase to include additional algorithms and allow for flexible configuration of algorithm implementation details such as parameter sharing. Finally, we open-source two environments for multi-agent research which focus on coordination under sparse rewards.

Lukas Schäfer, Logan Jones, Anssi Kanervisto et al. · microsoft-research

Video games have served as useful benchmarks for the decision-making community, but going beyond Atari games towards modern games has been prohibitively expensive for the vast majority of the research community. Prior work in modern video games typically relied on game-specific integration to obtain game features and enable online training, or on existing large datasets. An alternative approach is to train agents using imitation learning to play video games purely from images. However, this setting poses a fundamental question: which visual encoders obtain representations that retain information critical for decision making? To answer this question, we conduct a systematic study of imitation learning with publicly available pre-trained visual encoders compared to the typical task-specific end-to-end training approach in Minecraft, Counter-Strike: Global Offensive, and Minecraft Dungeons. Our results show that end-to-end training can be effective with comparably low-resolution images and only minutes of demonstrations, but significant improvements can be gained by utilising pre-trained encoders such as DINOv2 depending on the game. In addition to enabling effective decision making, we show that pre-trained encoders can make decision-making research in video games more accessible by significantly reducing the cost of training.

Hannah Markgraf, Shamburaj Sawant, Hanna Krasowski et al.

Projection-based safety filters, which modify unsafe actions by mapping them to the closest safe alternative, are widely used to enforce safety constraints in reinforcement learning (RL). Two integration strategies are commonly considered: Safe environment RL (SE-RL), where the safeguard is treated as part of the environment, and safe policy RL (SP-RL), where it is embedded within the policy through differentiable optimization layers. Despite their practical relevance in safety-critical settings, a formal understanding of their differences is lacking. In this work, we present a theoretical comparison of SE-RL and SP-RL. We identify a key distinction in how each approach is affected by action aliasing, a phenomenon in which multiple unsafe actions are projected to the same safe action, causing information loss in the policy gradients. In SE-RL, this effect is implicitly approximated by the critic, while in SP-RL, it manifests directly as rank-deficient Jacobians during backpropagation through the safeguard. Our contributions are threefold: (i) a unified formalization of SE-RL and SP-RL in the context of actor-critic algorithms, (ii) a theoretical analysis of their respective policy gradient estimates, highlighting the role of action aliasing, and (iii) a comparative study of mitigation strategies, including a novel penalty-based improvement for SP-RL that aligns with established SE-RL practices. Empirical results support our theoretical predictions, showing that action aliasing is more detrimental for SP-RL than for SE-RL. However, with appropriate improvement strategies, SP-RL can match or outperform improved SE-RL across a range of environments. These findings provide actionable insights for choosing and refining projection-based safe RL methods based on task characteristics.

Rujie Zhong, Duohan Zhang, Lukas Schäfer et al.

Reinforcement learning (RL) algorithms are often categorized as either on-policy or off-policy depending on whether they use data from a target policy of interest or from a different behavior policy. In this paper, we study a subtle distinction between on-policy data and on-policy sampling in the context of the RL sub-problem of policy evaluation. We observe that on-policy sampling may fail to match the expected distribution of on-policy data after observing only a finite number of trajectories and this failure hinders data-efficient policy evaluation. Towards improved data-efficiency, we show how non-i.i.d., off-policy sampling can produce data that more closely matches the expected on-policy data distribution and consequently increases the accuracy of the Monte Carlo estimator for policy evaluation. We introduce a method called Robust On-Policy Sampling and demonstrate theoretically and empirically that it produces data that converges faster to the expected on-policy distribution compared to on-policy sampling. Empirically, we show that this faster convergence leads to lower mean squared error policy value estimates.

Trevor McInroe, Lukas Schäfer, Stefano V. Albrecht

Deep reinforcement learning (RL) agents that exist in high-dimensional state spaces, such as those composed of images, have interconnected learning burdens. Agents must learn an action-selection policy that completes their given task, which requires them to learn a representation of the state space that discerns between useful and useless information. The reward function is the only supervised feedback that RL agents receive, which causes a representation learning bottleneck that can manifest in poor sample efficiency. We present $k$-Step Latent (KSL), a new representation learning method that enforces temporal consistency of representations via a self-supervised auxiliary task wherein agents learn to recurrently predict action-conditioned representations of the state space. The state encoder learned by KSL produces low-dimensional representations that make optimization of the RL task more sample efficient. Altogether, KSL produces state-of-the-art results in both data efficiency and asymptotic performance in the popular PlaNet benchmark suite. Our analyses show that KSL produces encoders that generalize better to new tasks unseen during training, and its representations are more strongly tied to reward, are more invariant to perturbations in the state space, and move more smoothly through the temporal axis of the RL problem than other methods such as DrQ, RAD, CURL, and SAC-AE.

Lukas Schäfer, Filippos Christianos, Josiah P. Hanna et al.

Intrinsic rewards can improve exploration in reinforcement learning, but the exploration process may suffer from instability caused by non-stationary reward shaping and strong dependency on hyperparameters. In this work, we introduce Decoupled RL (DeRL) as a general framework which trains separate policies for intrinsically-motivated exploration and exploitation. Such decoupling allows DeRL to leverage the benefits of intrinsic rewards for exploration while demonstrating improved robustness and sample efficiency. We evaluate DeRL algorithms in two sparse-reward environments with multiple types of intrinsic rewards. Our results show that DeRL is more robust to varying scale and rate of decay of intrinsic rewards and converges to the same evaluation returns than intrinsically-motivated baselines in fewer interactions. Lastly, we discuss the challenge of distribution shift and show that divergence constraint regularisers can successfully minimise instability caused by divergence of exploration and exploitation policies.

Filippos Christianos, Lukas Schäfer, Stefano V. Albrecht

Exploration in multi-agent reinforcement learning is a challenging problem, especially in environments with sparse rewards. We propose a general method for efficient exploration by sharing experience amongst agents. Our proposed algorithm, called Shared Experience Actor-Critic (SEAC), applies experience sharing in an actor-critic framework. We evaluate SEAC in a collection of sparse-reward multi-agent environments and find that it consistently outperforms two baselines and two state-of-the-art algorithms by learning in fewer steps and converging to higher returns. In some harder environments, experience sharing makes the difference between learning to solve the task and not learning at all.