Christopher Amato

Kevin Esslinger, Robert Platt, Christopher Amato

Real-world reinforcement learning tasks often involve some form of partial observability where the observations only give a partial or noisy view of the true state of the world. Such tasks typically require some form of memory, where the agent has access to multiple past observations, in order to perform well. One popular way to incorporate memory is by using a recurrent neural network to access the agent's history. However, recurrent neural networks in reinforcement learning are often fragile and difficult to train, susceptible to catastrophic forgetting and sometimes fail completely as a result. In this work, we propose Deep Transformer Q-Networks (DTQN), a novel architecture utilizing transformers and self-attention to encode an agent's history. DTQN is designed modularly, and we compare results against several modifications to our base model. Our experiments demonstrate the transformer can solve partially observable tasks faster and more stably than previous recurrent approaches.

Hai Nguyen, Andrea Baisero, Dian Wang et al.

Reinforcement learning in partially observable domains is challenging due to the lack of observable state information. Thankfully, learning offline in a simulator with such state information is often possible. In particular, we propose a method for partially observable reinforcement learning that uses a fully observable policy (which we call a state expert) during offline training to improve online performance. Based on Soft Actor-Critic (SAC), our agent balances performing actions similar to the state expert and getting high returns under partial observability. Our approach can leverage the fully-observable policy for exploration and parts of the domain that are fully observable while still being able to learn under partial observability. On six robotics domains, our method outperforms pure imitation, pure reinforcement learning, the sequential or parallel combination of both types, and a recent state-of-the-art method in the same setting. A successful policy transfer to a physical robot in a manipulation task from pixels shows our approach's practicality in learning interesting policies under partial observability.

Yuchen Xiao, Weihao Tan, Christopher Amato

Synchronizing decisions across multiple agents in realistic settings is problematic since it requires agents to wait for other agents to terminate and communicate about termination reliably. Ideally, agents should learn and execute asynchronously instead. Such asynchronous methods also allow temporally extended actions that can take different amounts of time based on the situation and action executed. Unfortunately, current policy gradient methods are not applicable in asynchronous settings, as they assume that agents synchronously reason about action selection at every time step. To allow asynchronous learning and decision-making, we formulate a set of asynchronous multi-agent actor-critic methods that allow agents to directly optimize asynchronous policies in three standard training paradigms: decentralized learning, centralized learning, and centralized training for decentralized execution. Empirical results (in simulation and hardware) in a variety of realistic domains demonstrate the superiority of our approaches in large multi-agent problems and validate the effectiveness of our algorithms for learning high-quality and asynchronous solutions.

Rohit Bokade, Xiaoning Jin, Christopher Amato

Traffic signal control (TSC) is a challenging problem within intelligent transportation systems and has been tackled using multi-agent reinforcement learning (MARL). While centralized approaches are often infeasible for large-scale TSC problems, decentralized approaches provide scalability but introduce new challenges, such as partial observability. Communication plays a critical role in decentralized MARL, as agents must learn to exchange information using messages to better understand the system and achieve effective coordination. Deep MARL has been used to enable inter-agent communication by learning communication protocols in a differentiable manner. However, many deep MARL communication frameworks proposed for TSC allow agents to communicate with all other agents at all times, which can add to the existing noise in the system and degrade overall performance. In this study, we propose a communication-based MARL framework for large-scale TSC. Our framework allows each agent to learn a communication policy that dictates "which" part of the message is sent "to whom". In essence, our framework enables agents to selectively choose the recipients of their messages and exchange variable length messages with them. This results in a decentralized and flexible communication mechanism in which agents can effectively use the communication channel only when necessary. We designed two networks, a synthetic $4 \times 4$ grid network and a real-world network based on the Pasubio neighborhood in Bologna. Our framework achieved the lowest network congestion compared to related methods, with agents utilizing $\sim 47-65 \%$ of the communication channel. Ablation studies further demonstrated the effectiveness of the communication policies learned within our framework.

Xueguang Lyu, Andrea Baisero, Yuchen Xiao et al.

Centralized Training for Decentralized Execution where agents are trained offline in a centralized fashion and execute online in a decentralized manner, has become a popular approach in Multi-Agent Reinforcement Learning (MARL). In particular, it has become popular to develop actor-critic methods that train decentralized actors with a centralized critic where the centralized critic is allowed access global information of the entire system, including the true system state. Such centralized critics are possible given offline information and are not used for online execution. While these methods perform well in a number of domains and have become a de facto standard in MARL, using a centralized critic in this context has yet to be sufficiently analyzed theoretically or empirically. In this paper, we therefore formally analyze centralized and decentralized critic approaches, and analyze the effect of using state-based critics in partially observable environments. We derive theories contrary to the common intuition: critic centralization is not strictly beneficial, and using state values can be harmful. We further prove that, in particular, state-based critics can introduce unexpected bias and variance compared to history-based critics. Finally, we demonstrate how the theory applies in practice by comparing different forms of critics on a wide range of common multi-agent benchmarks. The experiments show practical issues such as the difficulty of representation learning with partial observability, which highlights why the theoretical problems are often overlooked in the literature.

Enrico Marchesini, Christopher Amato

Deep Policy Gradient (PG) algorithms employ value networks to drive the learning of parameterized policies and reduce the variance of the gradient estimates. However, value function approximation gets stuck in local optima and struggles to fit the actual return, limiting the variance reduction efficacy and leading policies to sub-optimal performance. This paper focuses on improving value approximation and analyzing the effects on Deep PG primitives such as value prediction, variance reduction, and correlation of gradient estimates with the true gradient. To this end, we introduce a Value Function Search that employs a population of perturbed value networks to search for a better approximation. Our framework does not require additional environment interactions, gradient computations, or ensembles, providing a computationally inexpensive approach to enhance the supervised learning task on which value networks train. Crucially, we show that improving Deep PG primitives results in improved sample efficiency and policies with higher returns using common continuous control benchmark domains.

Enrico Marchesini, Luca Marzari, Alessandro Farinelli et al.

Cost functions are commonly employed in Safe Deep Reinforcement Learning (DRL). However, the cost is typically encoded as an indicator function due to the difficulty of quantifying the risk of policy decisions in the state space. Such an encoding requires the agent to visit numerous unsafe states to learn a cost-value function to drive the learning process toward safety. Hence, increasing the number of unsafe interactions and decreasing sample efficiency. In this paper, we investigate an alternative approach that uses domain knowledge to quantify the risk in the proximity of such states by defining a violation metric. This metric is computed by verifying task-level properties, shaped as input-output conditions, and it is used as a penalty to bias the policy away from unsafe states without learning an additional value function. We investigate the benefits of using the violation metric in standard Safe DRL benchmarks and robotic mapless navigation tasks. The navigation experiments bridge the gap between Safe DRL and robotics, introducing a framework that allows rapid testing on real robots. Our experiments show that policies trained with the violation penalty achieve higher performance over Safe DRL baselines and significantly reduce the number of visited unsafe states.

Brett Daley, Martha White, Christopher Amato et al.

Off-policy learning from multistep returns is crucial for sample-efficient reinforcement learning, but counteracting off-policy bias without exacerbating variance is challenging. Classically, off-policy bias is corrected in a per-decision manner: past temporal-difference errors are re-weighted by the instantaneous Importance Sampling (IS) ratio after each action via eligibility traces. Many off-policy algorithms rely on this mechanism, along with differing protocols for cutting the IS ratios to combat the variance of the IS estimator. Unfortunately, once a trace has been fully cut, the effect cannot be reversed. This has led to the development of credit-assignment strategies that account for multiple past experiences at a time. These trajectory-aware methods have not been extensively analyzed, and their theoretical justification remains uncertain. In this paper, we propose a multistep operator that can express both per-decision and trajectory-aware methods. We prove convergence conditions for our operator in the tabular setting, establishing the first guarantees for several existing methods as well as many new ones. Finally, we introduce Recency-Bounded Importance Sampling (RBIS), which leverages trajectory awareness to perform robustly across $λ$-values in an off-policy control task.

Hai Nguyen, Sammie Katt, Yuchen Xiao et al.

Robot learning is often difficult due to the expense of gathering data. The need for large amounts of data can, and should, be tackled with effective algorithms and leveraging expert information on robot dynamics. Bayesian reinforcement learning (BRL), thanks to its sample efficiency and ability to exploit prior knowledge, is uniquely positioned as such a solution method. Unfortunately, the application of BRL has been limited due to the difficulties of representing expert knowledge as well as solving the subsequent inference problem. This paper advances BRL for robotics by proposing a specialized framework for physical systems. In particular, we capture this knowledge in a factored representation, then demonstrate the posterior factorizes in a similar shape, and ultimately formalize the model in a Bayesian framework. We then introduce a sample-based online solution method, based on Monte-Carlo tree search and particle filtering, specialized to solve the resulting model. This approach can, for example, utilize typical low-level robot simulators and handle uncertainty over unknown dynamics of the environment. We empirically demonstrate its efficiency by performing on-robot learning in two human-robot interaction tasks with uncertainty about human behavior, achieving near-optimal performance after only a handful of real-world episodes. A video of learned policies is at https://youtu.be/H9xp60ngOes.

Hai Nguyen, Andrea Baisero, David Klee et al.

Incorporating inductive biases is a promising approach for tackling challenging robot learning domains with sample-efficient solutions. This paper identifies partially observable domains where symmetries can be a useful inductive bias for efficient learning. Specifically, by encoding the equivariance regarding specific group symmetries into the neural networks, our actor-critic reinforcement learning agents can reuse solutions in the past for related scenarios. Consequently, our equivariant agents outperform non-equivariant approaches significantly in terms of sample efficiency and final performance, demonstrated through experiments on a range of robotic tasks in simulation and real hardware.

Shuo Liu, Tianle Chen, Ryan Amiri et al.

Recent work has explored optimizing LLM collaboration through Multi-Agent Reinforcement Learning (MARL). However, most MARL fine-tuning approaches rely on predefined execution protocols, which often require centralized execution. Decentralized LLM collaboration is more appealing in practice, as agents can run inference in parallel with flexible deployments. Also, current approaches use Monte Carlo methods for fine-tuning, which suffer from high variance and thus require more samples to train effectively. Actor-critic methods are prevalent in MARL for dealing with these issues, so we developed Multi-Agent Actor-Critic (MAAC) methods to optimize decentralized LLM collaboration. In this paper, we analyze when and why these MAAC methods are beneficial. We propose 2 MAAC approaches, \textbf{CoLLM-CC} with a \textbf{C}entralized \textbf{C}ritic and \textbf{CoLLM-DC} with \textbf{D}ecentralized \textbf{C}ritics. Our experiments across writing, coding, and game-playing domains show that Monte Carlo methods and CoLLM-DC can achieve performance comparable to CoLLM-CC in short-horizon and dense-reward settings. However, they both underperform CoLLM-CC on long-horizon or sparse-reward tasks, where Monte Carlo methods require substantially more samples and CoLLM-DC struggles to converge. Our code is available at https://github.com/OpenMLRL/CoMLRL/releases/tag/v1.3.6.

Enrico Marchesini, Andrea Baisero, Rupali Bhati et al.

Value factorization is a popular paradigm for designing scalable multi-agent reinforcement learning algorithms. However, current factorization methods make choices without full justification that may limit their performance. For example, the theory in prior work uses stateless (i.e., history) functions, while the practical implementations use state information -- making the motivating theory a mismatch for the implementation. Also, methods have built off of previous approaches, inheriting their architectures without exploring other, potentially better ones. To address these concerns, we formally analyze the theory of using the state instead of the history in current methods -- reconnecting theory and practice. We then introduce DuelMIX, a factorization algorithm that learns distinct per-agent utility estimators to improve performance and achieve full expressiveness. Experiments on StarCraft II micromanagement and Box Pushing tasks demonstrate the benefits of our intuitions.

Christopher Amato

Multi-agent reinforcement learning (MARL) has exploded in popularity in recent years. Many approaches have been developed but they can be divided into three main types: centralized training and execution (CTE), centralized training for decentralized execution (CTDE), and Decentralized training and execution (DTE). CTDE methods are the most common as they can use centralized information during training but execute in a decentralized manner -- using only information available to that agent during execution. CTDE is the only paradigm that requires a separate training phase where any available information (e.g., other agent policies, underlying states) can be used. As a result, they can be more scalable than CTE methods, do not require communication during execution, and can often perform well. CTDE fits most naturally with the cooperative case, but can be potentially applied in competitive or mixed settings depending on what information is assumed to be observed. This text is an introduction to CTDE in cooperative MARL. It is meant to explain the setting, basic concepts, and common methods. It does not cover all work in CTDE MARL as the subarea is quite extensive. I have included work that I believe is important for understanding the main concepts in the subarea and apologize to those that I have omitted.

Chengguang Xu, Hieu T. Nguyen, Christopher Amato et al.

Navigating in unseen environments is crucial for mobile robots. Enhancing them with the ability to follow instructions in natural language will further improve navigation efficiency in unseen cases. However, state-of-the-art (SOTA) vision-and-language navigation (VLN) methods are mainly evaluated in simulation, neglecting the complex and noisy real world. Directly transferring SOTA navigation policies trained in simulation to the real world is challenging due to the visual domain gap and the absence of prior knowledge about unseen environments. In this work, we propose a novel navigation framework to address the VLN task in the real world. Utilizing the powerful foundation models, the proposed framework includes four key components: (1) an LLMs-based instruction parser that converts the language instruction into a sequence of pre-defined macro-action descriptions, (2) an online visual-language mapper that builds a real-time visual-language map to maintain a spatial and semantic understanding of the unseen environment, (3) a language indexing-based localizer that grounds each macro-action description into a waypoint location on the map, and (4) a DD-PPO-based local controller that predicts the action. We evaluate the proposed pipeline on an Interbotix LoCoBot WX250 in an unseen lab environment. Without any fine-tuning, our pipeline significantly outperforms the SOTA VLN baseline in the real world.

Chulabhaya Wijesundara, Andrea Baisero, Zhongheng Li et al.

Centralized training with decentralized execution (CTDE) is a standard framework for cooperative multi-agent policy-gradient reinforcement learning, allowing agents to learn from joint information while acting from local observations. Ratio-based trust-region methods such as Multi-Agent Proximal Policy Optimization (MAPPO) and Multi-Agent Simple Policy Optimization (MASPO) update decentralized actors using per-agent probability ratios weighted by joint advantage estimates. Teammate non-stationarity increases the variance of these advantages, which in turn increases the variance in the local ratio updates. This exposes two method-specific failure modes: MAPPO's additive clipping removes gradients for outlier samples and weakens recovery from policy drift, while MASPO's soft quadratic penalty can allow probability collapse. We introduce Multi-Agent Ratio Symmetry (MARS), a novel policy optimization objective that replaces these additive ratio-based trust-region mechanisms with a multiplicatively symmetric geometric barrier. MARS preserves corrective gradients while assigning unbounded cost as probability ratios approach zero. Across 47 tasks spanning eight multi-agent environments, including novel JAX benchmarks PaxMen and AeroJAX, MARS matches or exceeds MAPPO and MASPO in aggregate environment-level performance. Ablations show that these gains arise from the geometry of the symmetric barrier rather than from flexible trust-region boundaries alone.

Shuo Liu, Xinzichen Li, Christopher Amato

Robust embodied navigation relies on complementary sensory cues. However, high-quality and well-aligned multi-modal data is often difficult to obtain in practice. Training a monolithic model is also challenging as rich multi-modal inputs induce complex representations and substantially enlarge the policy space. Cross-modal collaboration among lightweight modality-specialized agents offers a scalable paradigm. It enables flexible deployment and parallel execution, while preserving the strength of each modality. In this paper, we propose \textbf{CRONA}, a Multi-Agent Reinforcement Learning (MARL) framework for \textbf{Cro}ss-Modal \textbf{Na}vigation. CRONA improves collaboration by leveraging control-relevant auxiliary beliefs and a centralized multi-modal critic with global state. Experiments on visual-acoustic navigation tasks show that multi-agent methods significantly improve performance and efficiency over single-agent baselines. We find that homogeneous collaboration with limited modalities is sufficient for short-range navigation under salient cues; heterogeneous collaboration among agents with complementary modalities is generally efficient and effective; and navigation in large, complex environments requires both richer multi-modal perception and increased model capacity.

Shuo Liu, Tianle Chen, Zeyu Liang et al.

A large amount of work has been done in Multi-Agent Systems (MAS) for modeling and solving problems with multiple interacting agents. However, most LLMs are pretrained independently and not specifically optimized for coordination. Existing LLM fine-tuning frameworks rely on individual rewards, which require complex reward designs for each agent to encourage collaboration. To address these challenges, we model LLM collaboration as a cooperative Multi-Agent Reinforcement Learning (MARL) problem. We develop a multi-agent, multi-turn algorithm, Multi-Agent Group Relative Policy Optimization (MAGRPO), to solve it, building on current RL approaches for LLMs as well as MARL techniques. Our experiments on LLM writing and coding collaboration demonstrate that fine-tuning MAS with MAGRPO enables agents to generate high-quality responses efficiently through effective cooperation. Our approach opens the door to using other MARL methods for LLMs and highlights the associated challenges.

Christopher Amato

Multi-agent reinforcement learning (MARL) has exploded in popularity in recent years. While numerous approaches have been developed, they can be broadly categorized into three main types: centralized training and execution (CTE), centralized training for decentralized execution (CTDE), and decentralized training and execution (DTE). CTE methods assume centralization during training and execution (e.g., with fast, free, and perfect communication) and have the most information during execution. CTDE methods are the most common, as they leverage centralized information during training while enabling decentralized execution -- using only information available to that agent during execution. Decentralized training and execution methods make the fewest assumptions and are often simple to implement. This text is an introduction to cooperative MARL -- MARL in which all agents share a single, joint reward. It is meant to explain the setting, basic concepts, and common methods for the CTE, CTDE, and DTE settings. It does not cover all work in cooperative MARL as the area is quite extensive. I have included work that I believe is important for understanding the main concepts in the area and apologize to those that I have omitted. Topics include simple applications of single-agent methods to CTE as well as some more scalable methods that exploit the multi-agent structure, independent Q-learning and policy gradient methods and their extensions, as well as value function factorization methods including the well-known VDN, QMIX, and QPLEX approaches, and centralized critic methods including MADDPG, COMA, and MAPPO. I also discuss common misconceptions, the relationship between different approaches, and some open questions.

Ayhan Alp Aydeniz, Enrico Marchesini, Robert Loftin et al.

Many real-world multiagent learning problems involve safety concerns. In these setups, typical safe reinforcement learning algorithms constrain agents' behavior, limiting exploration -- a crucial component for discovering effective cooperative multiagent behaviors. Moreover, the multiagent literature typically models individual constraints for each agent and has yet to investigate the benefits of using joint team constraints. In this work, we analyze these team constraints from a theoretical and practical perspective and propose entropic exploration for constrained multiagent reinforcement learning (E2C) to address the exploration issue. E2C leverages observation entropy maximization to incentivize exploration and facilitate learning safe and effective cooperative behaviors. Experiments across increasingly complex domains show that E2C agents match or surpass common unconstrained and constrained baselines in task performance while reducing unsafe behaviors by up to $50\%$.

Ethan Rathbun, Wo Wei Lin, Alina Oprea et al.

Simulated environments are a key piece in the success of Reinforcement Learning (RL), allowing practitioners and researchers to train decision making agents without running expensive experiments on real hardware. Simulators remain a security blind spot, however, enabling adversarial developers to alter the dynamics of their released simulators for malicious purposes. Therefore, in this work we highlight a novel threat, demonstrating how simulator dynamics can be exploited to stealthily implant action-level backdoors into RL agents. The backdoor then allows an adversary to reliably activate targeted actions in an agent upon observing a predefined ``trigger'', leading to potentially dangerous consequences. Traditional backdoor attacks are limited in their strong threat models, assuming the adversary has near full control over an agent's training pipeline, enabling them to both alter and observe agent's rewards. As these assumptions are infeasible to implement within a simulator, we propose a new attack ``Daze'' which is able to reliably and stealthily implant backdoors into RL agents trained for real world tasks without altering or even observing their rewards. We provide formal proof of Daze's effectiveness in guaranteeing attack success across general RL tasks along with extensive empirical evaluations on both discrete and continuous action space domains. We additionally provide the first example of RL backdoor attacks transferring to real, robotic hardware. These developments motivate further research into securing all components of the RL training pipeline to prevent malicious attacks.

Andrea Baisero, Rupali Bhati, Shuo Liu et al.

Value function decomposition methods for cooperative multi-agent reinforcement learning compose joint values from individual per-agent utilities, and train them using a joint objective. To ensure that the action selection process between individual utilities and joint values remains consistent, it is imperative for the composition to satisfy the individual-global max (IGM) property. Although satisfying IGM itself is straightforward, most existing methods (e.g., VDN, QMIX) have limited representation capabilities and are unable to represent the full class of IGM values, and the one exception that has no such limitation (QPLEX) is unnecessarily complex. In this work, we present a simple formulation of the full class of IGM values that naturally leads to the derivation of QFIX, a novel family of value function decomposition models that expand the representation capabilities of prior models by means of a thin "fixing" layer. We derive multiple variants of QFIX, and implement three variants in two well-known multi-agent frameworks. We perform an empirical evaluation on multiple SMACv2 and Overcooked environments, which confirms that QFIX (i) succeeds in enhancing the performance of prior methods, (ii) learns more stably and performs better than its main competitor QPLEX, and (iii) achieves this while employing the simplest and smallest mixing models.

Ethan Rathbun, Alina Oprea, Christopher Amato

Recent works have demonstrated the vulnerability of Deep Reinforcement Learning (DRL) algorithms against training-time, backdoor poisoning attacks. The objectives of these attacks are twofold: induce pre-determined, adversarial behavior in the agent upon observing a fixed trigger during deployment while allowing the agent to solve its intended task during training. Prior attacks assume arbitrary control over the agent's rewards, inducing values far outside the environment's natural constraints. This results in brittle attacks that fail once the proper reward constraints are enforced. Thus, in this work we propose a new class of backdoor attacks against DRL which are the first to achieve state of the art performance under strict reward constraints. These "inception" attacks manipulate the agent's training data -- inserting the trigger into prior observations and replacing high return actions with those of the targeted adversarial behavior. We formally define these attacks and prove they achieve both adversarial objectives against arbitrary Markov Decision Processes (MDP). Using this framework we devise an online inception attack which achieves an 100\% attack success rate on multiple environments under constrained rewards while minimally impacting the agent's task performance.

Sammie Katt, Hai Nguyen, Frans A. Oliehoek et al.

While reinforcement learning (RL) has made great advances in scalability, exploration and partial observability are still active research topics. In contrast, Bayesian RL (BRL) provides a principled answer to both state estimation and the exploration-exploitation trade-off, but struggles to scale. To tackle this challenge, BRL frameworks with various prior assumptions have been proposed, with varied success. This work presents a representation-agnostic formulation of BRL under partially observability, unifying the previous models under one theoretical umbrella. To demonstrate its practical significance we also propose a novel derivation, Bayes-Adaptive Deep Dropout rl (BADDr), based on dropout networks. Under this parameterization, in contrast to previous work, the belief over the state and dynamics is a more scalable inference problem. We choose actions through Monte-Carlo tree search and empirically show that our method is competitive with state-of-the-art BRL methods on small domains while being able to solve much larger ones.

Xueguang Lyu, Andrea Baisero, Yuchen Xiao et al.

Centralized Training for Decentralized Execution, where training is done in a centralized offline fashion, has become a popular solution paradigm in Multi-Agent Reinforcement Learning. Many such methods take the form of actor-critic with state-based critics, since centralized training allows access to the true system state, which can be useful during training despite not being available at execution time. State-based critics have become a common empirical choice, albeit one which has had limited theoretical justification or analysis. In this paper, we show that state-based critics can introduce bias in the policy gradient estimates, potentially undermining the asymptotic guarantees of the algorithm. We also show that, even if the state-based critics do not introduce any bias, they can still result in a larger gradient variance, contrary to the common intuition. Finally, we show the effects of the theories in practice by comparing different forms of centralized critics on a wide range of common benchmarks, and detail how various environmental properties are related to the effectiveness of different types of critics.

Brett Daley, Christopher Amato

Off-policy learning from multistep returns is crucial for sample-efficient reinforcement learning, particularly in the experience replay setting now commonly used with deep neural networks. Classically, off-policy estimation bias is corrected in a per-decision manner: past temporal-difference errors are re-weighted by the instantaneous Importance Sampling (IS) ratio (via eligibility traces) after each action. Many important off-policy algorithms such as Tree Backup and Retrace rely on this mechanism along with differing protocols for truncating ("cutting") the ratios ("traces") to counteract the excessive variance of the IS estimator. Unfortunately, cutting traces on a per-decision basis is not necessarily efficient; once a trace has been cut according to local information, the effect cannot be reversed later, potentially resulting in the premature truncation of estimated returns and slower learning. In the interest of motivating efficient off-policy algorithms, we propose a multistep operator that permits arbitrary past-dependent traces. We prove that our operator is convergent for policy evaluation, and for optimal control when targeting greedy-in-the-limit policies. Our theorems establish the first convergence guarantees for many existing algorithms including Truncated IS, Non-Markov Retrace, and history-dependent TD($λ$). Our theoretical results also provide guidance for the development of new algorithms that jointly consider multiple past decisions for better credit assignment and faster learning.

Brett Daley, Christopher Amato

Return caching is a recent strategy that enables efficient minibatch training with multistep estimators (e.g. the λ-return) for deep reinforcement learning. By precomputing return estimates in sequential batches and then storing the results in an auxiliary data structure for later sampling, the average computation spent per estimate can be greatly reduced. Still, the efficiency of return caching could be improved, particularly with regard to its large memory usage and repetitive data copies. We propose a new data structure, the Virtual Replay Cache (VRC), to address these shortcomings. When learning to play Atari 2600 games, the VRC nearly eliminates DQN(λ)'s cache memory footprint and slightly reduces the total training time on our hardware.

Brett Daley, Christopher Amato

Deep Q-Network (DQN) marked a major milestone for reinforcement learning, demonstrating for the first time that human-level control policies could be learned directly from raw visual inputs via reward maximization. Even years after its introduction, DQN remains highly relevant to the research community since many of its innovations have been adopted by successor methods. Nevertheless, despite significant hardware advances in the interim, DQN's original Atari 2600 experiments remain costly to replicate in full. This poses an immense barrier to researchers who cannot afford state-of-the-art hardware or lack access to large-scale cloud computing resources. To facilitate improved access to deep reinforcement learning research, we introduce a DQN implementation that leverages a novel concurrent and synchronized execution framework designed to maximally utilize a heterogeneous CPU-GPU desktop system. With just one NVIDIA GeForce GTX 1080 GPU, our implementation reduces the training time of a 200-million-frame Atari experiment from 25 hours to just 9 hours. The ideas introduced in our paper should be generalizable to a large number of off-policy deep reinforcement learning methods.

Yuchen Xiao, Xueguang Lyu, Christopher Amato

Policy gradient methods have become popular in multi-agent reinforcement learning, but they suffer from high variance due to the presence of environmental stochasticity and exploring agents (i.e., non-stationarity), which is potentially worsened by the difficulty in credit assignment. As a result, there is a need for a method that is not only capable of efficiently solving the above two problems but also robust enough to solve a variety of tasks. To this end, we propose a new multi-agent policy gradient method, called Robust Local Advantage (ROLA) Actor-Critic. ROLA allows each agent to learn an individual action-value function as a local critic as well as ameliorating environment non-stationarity via a novel centralized training approach based on a centralized critic. By using this local critic, each agent calculates a baseline to reduce variance on its policy gradient estimation, which results in an expected advantage action-value over other agents' choices that implicitly improves credit assignment. We evaluate ROLA across diverse benchmarks and show its robustness and effectiveness over a number of state-of-the-art multi-agent policy gradient algorithms.

Brett Daley, Christopher Amato

Adam is an adaptive gradient method that has experienced widespread adoption due to its fast and reliable training performance. Recent approaches have not offered significant improvement over Adam, often because they do not innovate upon one of its core features: normalization by the root mean square (RMS) of recent gradients. However, as noted by Kingma and Ba (2015), any number of $L^p$ normalizations are possible, with the RMS corresponding to the specific case of $p=2$. In our work, we theoretically and empirically characterize the influence of different $L^p$ norms on adaptive gradient methods for the first time. We show mathematically how the choice of $p$ influences the size of the steps taken, while leaving other desirable properties unaffected. We evaluate Adam with various $L^p$ norms on a suite of deep learning benchmarks, and find that $p > 2$ consistently leads to improved learning speed and final performance. The choices of $p=3$ or $p=6$ also match or outperform state-of-the-art methods in all of our experiments.

Andrea Baisero, Christopher Amato

Predictive state representations (PSRs) are models of controlled non-Markov observation sequences which exhibit the same generative process governing POMDP observations without relying on an underlying latent state. In that respect, a PSR is indistinguishable from the corresponding POMDP. However, PSRs notoriously ignore the notion of rewards, which undermines the general utility of PSR models for control, planning, or reinforcement learning. Therefore, we describe a sufficient and necessary accuracy condition which determines whether a PSR is able to accurately model POMDP rewards, we show that rewards can be approximated even when the accuracy condition is not satisfied, and we find that a non-trivial number of POMDPs taken from a well-known third-party repository do not satisfy the accuracy condition. We propose reward-predictive state representations (R-PSRs), a generalization of PSRs which accurately models both observations and rewards, and develop value iteration for R-PSRs. We show that there is a mismatch between optimal POMDP policies and the optimal PSR policies derived from approximate rewards. On the other hand, optimal R-PSR policies perfectly match optimal POMDP policies, reconfirming R-PSRs as accurate state-less generative models of observations and rewards.

Chengguang Xu, Christopher Amato, Lawson L. S. Wong

In robot navigation, generalizing quickly to unseen environments is essential. Hierarchical methods inspired by human navigation have been proposed, typically consisting of a high-level landmark proposer and a low-level controller. However, these methods either require precise high-level information to be given in advance or need to construct such guidance from extensive interaction with the environment. In this work, we propose an approach that leverages a rough 2-D map of the environment to navigate in novel environments without requiring further learning. In particular, we introduce a dynamic topological map that can be initialized from the rough 2-D map along with a high-level planning approach for proposing reachable 2-D map patches of the intermediate landmarks between the start and goal locations. To use proposed 2-D patches, we train a deep generative model to generate intermediate landmarks in observation space which are used as subgoals by low-level goal-conditioned reinforcement learning. Importantly, because the low-level controller is only trained with local behaviors (e.g. go across the intersection, turn left at a corner) on existing environments, this framework allows us to generalize to novel environments given only a rough 2-D map, without requiring further learning. Experimental results demonstrate the effectiveness of the proposed framework in both seen and novel environments.

Andrea Baisero, Christopher Amato

In partially observable reinforcement learning, offline training gives access to latent information which is not available during online training and/or execution, such as the system state. Asymmetric actor-critic methods exploit such information by training a history-based policy via a state-based critic. However, many asymmetric methods lack theoretical foundation, and are only evaluated on limited domains. We examine the theory of asymmetric actor-critic methods which use state-based critics, and expose fundamental issues which undermine the validity of a common variant, and limit its ability to address partial observability. We propose an unbiased asymmetric actor-critic variant which is able to exploit state information while remaining theoretically sound, maintaining the validity of the policy gradient theorem, and introducing no bias and relatively low variance into the training process. An empirical evaluation performed on domains which exhibit significant partial observability confirms our analysis, demonstrating that unbiased asymmetric actor-critic converges to better policies and/or faster than symmetric and biased asymmetric baselines.

Mohammadreza Sharif, Deniz Erdogmus, Christopher Amato et al.

State-of-the-art human-in-the-loop robot grasping is hugely suffered by Electromyography (EMG) inference robustness issues. As a workaround, researchers have been looking into integrating EMG with other signals, often in an ad hoc manner. In this paper, we are presenting a method for end-to-end training of a policy for human-in-the-loop robot grasping on real reaching trajectories. For this purpose we use Reinforcement Learning (RL) and Imitation Learning (IL) in DEXTRON (DEXTerity enviRONment), a stochastic simulation environment with real human trajectories that are augmented and selected using a Monte Carlo (MC) simulation method. We also offer a success model which once trained on the expert policy data and the RL policy roll-out transitions, can provide transparency to how the deep policy works and when it is probably going to fail.

Roi Yehoshua, Juan Heredia-Juesas, Yushu Wu et al.

Targets search and detection encompasses a variety of decision problems such as coverage, surveillance, search, observing and pursuit-evasion along with others. In this paper we develop a multi-agent deep reinforcement learning (MADRL) method to coordinate a group of aerial vehicles (drones) for the purpose of locating a set of static targets in an unknown area. To that end, we have designed a realistic drone simulator that replicates the dynamics and perturbations of a real experiment, including statistical inferences taken from experimental data for its modeling. Our reinforcement learning method, which utilized this simulator for training, was able to find near-optimal policies for the drones. In contrast to other state-of-the-art MADRL methods, our method is fully decentralized during both learning and execution, can handle high-dimensional and continuous observation spaces, and does not require tuning of additional hyperparameters.

Brett Daley, Cameron Hickert, Christopher Amato

Deep Reinforcement Learning (RL) methods rely on experience replay to approximate the minibatched supervised learning setting; however, unlike supervised learning where access to lots of training data is crucial to generalization, replay-based deep RL appears to struggle in the presence of extraneous data. Recent works have shown that the performance of Deep Q-Network (DQN) degrades when its replay memory becomes too large. This suggests that outdated experiences somehow impact the performance of deep RL, which should not be the case for off-policy methods like DQN. Consequently, we re-examine the motivation for sampling uniformly over a replay memory, and find that it may be flawed when using function approximation. We show that -- despite conventional wisdom -- sampling from the uniform distribution does not yield uncorrelated training samples and therefore biases gradients during training. Our theory prescribes a special non-uniform distribution to cancel this effect, and we propose a stratified sampling scheme to efficiently implement it.

Xueguang Lyu, Yuchen Xiao, Brett Daley et al.

Centralized Training for Decentralized Execution, where agents are trained offline using centralized information but execute in a decentralized manner online, has gained popularity in the multi-agent reinforcement learning community. In particular, actor-critic methods with a centralized critic and decentralized actors are a common instance of this idea. However, the implications of using a centralized critic in this context are not fully discussed and understood even though it is the standard choice of many algorithms. We therefore formally analyze centralized and decentralized critic approaches, providing a deeper understanding of the implications of critic choice. Because our theory makes unrealistic assumptions, we also empirically compare the centralized and decentralized critic methods over a wide set of environments to validate our theories and to provide practical advice. We show that there exist misconceptions regarding centralized critics in the current literature and show that the centralized critic design is not strictly beneficial, but rather both centralized and decentralized critics have different pros and cons that should be taken into account by algorithm designers.

Ingy Elsayed-Aly, Suda Bharadwaj, Christopher Amato et al.

Multi-agent reinforcement learning (MARL) has been increasingly used in a wide range of safety-critical applications, which require guaranteed safety (e.g., no unsafe states are ever visited) during the learning process.Unfortunately, current MARL methods do not have safety guarantees. Therefore, we present two shielding approaches for safe MARL. In centralized shielding, we synthesize a single shield to monitor all agents' joint actions and correct any unsafe action if necessary. In factored shielding, we synthesize multiple shields based on a factorization of the joint state space observed by all agents; the set of shields monitors agents concurrently and each shield is only responsible for a subset of agents at each step.Experimental results show that both approaches can guarantee the safety of agents during learning without compromising the quality of learned policies; moreover, factored shielding is more scalable in the number of agents than centralized shielding.

Hai Nguyen, Brett Daley, Xinchao Song et al.

Many important robotics problems are partially observable in the sense that a single visual or force-feedback measurement is insufficient to reconstruct the state. Standard approaches involve learning a policy over beliefs or observation-action histories. However, both of these have drawbacks; it is expensive to track the belief online, and it is hard to learn policies directly over histories. We propose a method for policy learning under partial observability called the Belief-Grounded Network (BGN) in which an auxiliary belief-reconstruction loss incentivizes a neural network to concisely summarize its input history. Since the resulting policy is a function of the history rather than the belief, it can be executed easily at runtime. We compare BGN against several baselines on classic benchmark tasks as well as three novel robotic touch-sensing tasks. BGN outperforms all other tested methods and its learned policies work well when transferred onto a physical robot.

Brett Daley, Christopher Amato

Many popular adaptive gradient methods such as Adam and RMSProp rely on an exponential moving average (EMA) to normalize their stepsizes. While the EMA makes these methods highly responsive to new gradient information, recent research has shown that it also causes divergence on at least one convex optimization problem. We propose a novel method called Expectigrad, which adjusts stepsizes according to a per-component unweighted mean of all historical gradients and computes a bias-corrected momentum term jointly between the numerator and denominator. We prove that Expectigrad cannot diverge on every instance of the optimization problem known to cause Adam to diverge. We also establish a regret bound in the general stochastic nonconvex setting that suggests Expectigrad is less susceptible to gradient variance than existing methods are. Testing Expectigrad on several high-dimensional machine learning tasks, we find it often performs favorably to state-of-the-art methods with little hyperparameter tuning.

Yuchen Xiao, Joshua Hoffman, Christopher Amato

In real-world multi-robot systems, performing high-quality, collaborative behaviors requires robots to asynchronously reason about high-level action selection at varying time durations. Macro-Action Decentralized Partially Observable Markov Decision Processes (MacDec-POMDPs) provide a general framework for asynchronous decision making under uncertainty in fully cooperative multi-agent tasks. However, multi-agent deep reinforcement learning methods have only been developed for (synchronous) primitive-action problems. This paper proposes two Deep Q-Network (DQN) based methods for learning decentralized and centralized macro-action-value functions with novel macro-action trajectory replay buffers introduced for each case. Evaluations on benchmark problems and a larger domain demonstrate the advantage of learning with macro-actions over primitive-actions and the scalability of our approaches.

Christopher Amato, Andrea Baisero

To coordinate with other systems, agents must be able to determine what the systems are currently doing and predict what they will be doing in the future---plan and goal recognition. There are many methods for plan and goal recognition, but they assume a passive observer that continually monitors the target system. Real-world domains, where information gathering has a cost (e.g., moving a camera or a robot, or time taken away from another task), will often require a more active observer. We propose to combine goal recognition with other observer tasks in order to obtain \emph{active goal recognition} (AGR). We discuss this problem and provide a model and preliminary experimental results for one form of this composite problem. As expected, the results show that optimal behavior in AGR problems balance information gathering with other actions (e.g., task completion) such as to achieve all tasks jointly and efficiently. We hope that our formulation opens the door for extensive further research on this interesting and realistic problem.

Yuchen Xiao, Joshua Hoffman, Tian Xia et al.

In many real-world multi-robot tasks, high-quality solutions often require a team of robots to perform asynchronous actions under decentralized control. Decentralized multi-agent reinforcement learning methods have difficulty learning decentralized policies because of the environment appearing to be non-stationary due to other agents also learning at the same time. In this paper, we address this challenge by proposing a macro-action-based decentralized multi-agent double deep recurrent Q-net (MacDec-MADDRQN) which trains each decentralized Q-net using a centralized Q-net for action selection. A generalized version of MacDec-MADDRQN with two separate training environments, called Parallel-MacDec-MADDRQN, is also presented to leverage either centralized or decentralized exploration. The advantages and the practical nature of our methods are demonstrated by achieving near-centralized results in simulation and having real robots accomplish a warehouse tool delivery task in an efficient way.

Xueguang Lyu, Christopher Amato

When multiple agents learn in a decentralized manner, the environment appears non-stationary from the perspective of an individual agent due to the exploration and learning of the other agents. Recently proposed deep multi-agent reinforcement learning methods have tried to mitigate this non-stationarity by attempting to determine which samples are from other agent exploration or suboptimality and take them less into account during learning. Based on the same philosophy, this paper introduces a decentralized quantile estimator, which aims to improve performance by distinguishing non-stationary samples based on the likelihood of returns. In particular, each agent considers the likelihood that other agent exploration and policy changes are occurring, essentially utilizing the agent's own estimations to weigh the learning rate that should be applied towards the given samples. We introduce a formal method of calculating differences of our return distribution representations and methods for utilizing it to guide updates. We also explore the effect of risk-seeking strategies for adjusting learning over time and propose adaptive risk distortion functions which guides risk sensitivity. Our experiments, on traditional benchmarks and new domains, show our methods are more stable, sample efficient and more likely to converge to a joint optimal policy than previous methods.

Sammie Katt, Frans Oliehoek, Christopher Amato

Bayesian approaches provide a principled solution to the exploration-exploitation trade-off in Reinforcement Learning. Typical approaches, however, either assume a fully observable environment or scale poorly. This work introduces the Factored Bayes-Adaptive POMDP model, a framework that is able to exploit the underlying structure while learning the dynamics in partially observable systems. We also present a belief tracking method to approximate the joint posterior over state and model variables, and an adaptation of the Monte-Carlo Tree Search solution method, which together are capable of solving the underlying problem near-optimally. Our method is able to learn efficiently given a known factorization or also learn the factorization and the model parameters at the same time. We demonstrate that this approach is able to outperform current methods and tackle problems that were previously infeasible.

Brett Daley, Christopher Amato

Modern deep reinforcement learning methods have departed from the incremental learning required for eligibility traces, rendering the implementation of the $λ$-return difficult in this context. In particular, off-policy methods that utilize experience replay remain problematic because their random sampling of minibatches is not conducive to the efficient calculation of $λ$-returns. Yet replay-based methods are often the most sample efficient, and incorporating $λ$-returns into them is a viable way to achieve new state-of-the-art performance. Towards this, we propose the first method to enable practical use of $λ$-returns in arbitrary replay-based methods without relying on other forms of decorrelation such as asynchronous gradient updates. By promoting short sequences of past transitions into a small cache within the replay memory, adjacent $λ$-returns can be efficiently precomputed by sharing Q-values. Computation is not wasted on experiences that are never sampled, and stored $λ$-returns behave as stable temporal-difference (TD) targets that replace the target network. Additionally, our method grants the unique ability to observe TD errors prior to sampling; for the first time, transitions can be prioritized by their true significance rather than by a proxy to it. Furthermore, we propose the novel use of the TD error to dynamically select $λ$-values that facilitate faster learning. We show that these innovations can enhance the performance of DQN when playing Atari 2600 games, even under partial observability. While our work specifically focuses on $λ$-returns, these ideas are applicable to any multi-step return estimator.

Sammie Katt, Frans A. Oliehoek, Christopher Amato

The POMDP is a powerful framework for reasoning under outcome and information uncertainty, but constructing an accurate POMDP model is difficult. Bayes-Adaptive Partially Observable Markov Decision Processes (BA-POMDPs) extend POMDPs to allow the model to be learned during execution. BA-POMDPs are a Bayesian RL approach that, in principle, allows for an optimal trade-off between exploitation and exploration. Unfortunately, BA-POMDPs are currently impractical to solve for any non-trivial domain. In this paper, we extend the Monte-Carlo Tree Search method POMCP to BA-POMDPs and show that the resulting method, which we call BA-POMCP, is able to tackle problems that previous solution methods have been unable to solve. Additionally, we introduce several techniques that exploit the BA-POMDP structure to improve the efficiency of BA-POMCP along with proof of their convergence.

Shayegan Omidshafiei, Dong-Ki Kim, Miao Liu et al.

Collective human knowledge has clearly benefited from the fact that innovations by individuals are taught to others through communication. Similar to human social groups, agents in distributed learning systems would likely benefit from communication to share knowledge and teach skills. The problem of teaching to improve agent learning has been investigated by prior works, but these approaches make assumptions that prevent application of teaching to general multiagent problems, or require domain expertise for problems they can apply to. This learning to teach problem has inherent complexities related to measuring long-term impacts of teaching that compound the standard multiagent coordination challenges. In contrast to existing works, this paper presents the first general framework and algorithm for intelligent agents to learn to teach in a multiagent environment. Our algorithm, Learning to Coordinate and Teach Reinforcement (LeCTR), addresses peer-to-peer teaching in cooperative multiagent reinforcement learning. Each agent in our approach learns both when and what to advise, then uses the received advice to improve local learning. Importantly, these roles are not fixed; these agents learn to assume the role of student and/or teacher at the appropriate moments, requesting and providing advice in order to improve teamwide performance and learning. Empirical comparisons against state-of-the-art teaching methods show that our teaching agents not only learn significantly faster, but also learn to coordinate in tasks where existing methods fail.

Trong Nghia Hoang, Yuchen Xiao, Kavinayan Sivakumar et al.

A key challenge in multi-robot and multi-agent systems is generating solutions that are robust to other self-interested or even adversarial parties who actively try to prevent the agents from achieving their goals. The practicality of existing works addressing this challenge is limited to only small-scale synchronous decision-making scenarios or a single agent planning its best response against a single adversary with fixed, procedurally characterized strategies. In contrast this paper considers a more realistic class of problems where a team of asynchronous agents with limited observation and communication capabilities need to compete against multiple strategic adversaries with changing strategies. This problem necessitates agents that can coordinate to detect changes in adversary strategies and plan the best response accordingly. Our approach first optimizes a set of stratagems that represent these best responses. These optimized stratagems are then integrated into a unified policy that can detect and respond when the adversaries change their strategies. The near-optimality of the proposed framework is established theoretically as well as demonstrated empirically in simulation and hardware.

Miao Liu, Kavinayan Sivakumar, Shayegan Omidshafiei et al.

This paper presents a data-driven approach for multi-robot coordination in partially-observable domains based on Decentralized Partially Observable Markov Decision Processes (Dec-POMDPs) and macro-actions (MAs). Dec-POMDPs provide a general framework for cooperative sequential decision making under uncertainty and MAs allow temporally extended and asynchronous action execution. To date, most methods assume the underlying Dec-POMDP model is known a priori or a full simulator is available during planning time. Previous methods which aim to address these issues suffer from local optimality and sensitivity to initial conditions. Additionally, few hardware demonstrations involving a large team of heterogeneous robots and with long planning horizons exist. This work addresses these gaps by proposing an iterative sampling based Expectation-Maximization algorithm (iSEM) to learn polices using only trajectory data containing observations, MAs, and rewards. Our experiments show the algorithm is able to achieve better solution quality than the state-of-the-art learning-based methods. We implement two variants of multi-robot Search and Rescue (SAR) domains (with and without obstacles) on hardware to demonstrate the learned policies can effectively control a team of distributed robots to cooperate in a partially observable stochastic environment.

Shayegan Omidshafiei, Jason Pazis, Christopher Amato et al.

Many real-world tasks involve multiple agents with partial observability and limited communication. Learning is challenging in these settings due to local viewpoints of agents, which perceive the world as non-stationary due to concurrently-exploring teammates. Approaches that learn specialized policies for individual tasks face problems when applied to the real world: not only do agents have to learn and store distinct policies for each task, but in practice identities of tasks are often non-observable, making these approaches inapplicable. This paper formalizes and addresses the problem of multi-task multi-agent reinforcement learning under partial observability. We introduce a decentralized single-task learning approach that is robust to concurrent interactions of teammates, and present an approach for distilling single-task policies into a unified policy that performs well across multiple related tasks, without explicit provision of task identity.