Previous deep multi-agent reinforcement learning (MARL) algorithms have achieved impressive results, typically in homogeneous scenarios. However, heterogeneous scenarios are also very common and usually harder to solve. In this paper, we mainly discuss cooperative heterogeneous MARL problems in Starcraft Multi-Agent Challenges (SMAC) environment. We firstly define and describe the heterogeneous problems in SMAC. In order to comprehensively reveal and study the problem, we make new maps added to the original SMAC maps. We find that baseline algorithms fail to perform well in those heterogeneous maps. To address this issue, we propose the Grouped Individual-Global-Max Consistency (GIGM) and a novel MARL algorithm, Grouped Hybrid Q Learning (GHQ). GHQ separates agents into several groups and keeps individual parameters for each group, along with a novel hybrid structure for factorization. To enhance coordination between groups, we maximize the Inter-group Mutual Information (IGMI) between groups' trajectories. Experiments on original and new heterogeneous maps show the fabulous performance of GHQ compared to other state-of-the-art algorithms.
Recent methods for imitation learning directly learn a $Q$-function using an implicit reward formulation rather than an explicit reward function. However, these methods generally require implicit reward regularization to improve stability and often mistreat absorbing states. Previous works show that a squared norm regularization on the implicit reward function is effective, but do not provide a theoretical analysis of the resulting properties of the algorithms. In this work, we show that using this regularizer under a mixture distribution of the policy and the expert provides a particularly illuminating perspective: the original objective can be understood as squared Bellman error minimization, and the corresponding optimization problem minimizes a bounded $\chi^2$-Divergence between the expert and the mixture distribution. This perspective allows us to address instabilities and properly treat absorbing states. We show that our method, Least Squares Inverse Q-Learning (LS-IQ), outperforms state-of-the-art algorithms, particularly in environments with absorbing states. Finally, we propose to use an inverse dynamics model to learn from observations only. Using this approach, we retain performance in settings where no expert actions are available.
Text-based games are a popular testbed for language-based reinforcement learning (RL). In previous work, deep Q-learning is commonly used as the learning agent. Q-learning algorithms are challenging to apply to complex real-world domains due to, for example, their instability in training. Therefore, in this paper, we adapt the soft-actor-critic (SAC) algorithm to the text-based environment. To deal with sparse extrinsic rewards from the environment, we combine it with a potential-based reward shaping technique to provide more informative (dense) reward signals to the RL agent. We apply our method to play difficult text-based games. The SAC method achieves higher scores than the Q-learning methods on many games with only half the number of training steps. This shows that it is well-suited for text-based games. Moreover, we show that the reward shaping technique helps the agent to learn the policy faster and achieve higher scores. In particular, we consider a dynamically learned value function as a potential function for shaping the learner's original sparse reward signals.
Multi-agent reinforcement learning typically suffers from the problem of sample inefficiency, where learning suitable policies involves the use of many data samples. Learning from external demonstrators is a possible solution that mitigates this problem. However, most prior approaches in this area assume the presence of a single demonstrator. Leveraging multiple knowledge sources (i.e., advisors) with expertise in distinct aspects of the environment could substantially speed up learning in complex environments. This paper considers the problem of simultaneously learning from multiple independent advisors in multi-agent reinforcement learning. The approach leverages a two-level Q-learning architecture, and extends this framework from single-agent to multi-agent settings. We provide principled algorithms that incorporate a set of advisors by both evaluating the advisors at each state and subsequently using the advisors to guide action selection. We also provide theoretical convergence and sample complexity guarantees. Experimentally, we validate our approach in three different test-beds and show that our algorithms give better performances than baselines, can effectively integrate the combined expertise of different advisors, and learn to ignore bad advice.
Coordination is one of the most difficult aspects of multi-agent reinforcement learning (MARL). One reason is that agents normally choose their actions independently of one another. In order to see coordination strategies emerging from the combination of independent policies, the recent research has focused on the use of a centralized function (CF) that learns each agent's contribution to the team reward. However, the structure in which the environment is presented to the agents and to the CF is typically overlooked. We have observed that the features used to describe the coordination problem can be represented as vertex features of a latent graph structure. Here, we present TransfQMix, a new approach that uses transformers to leverage this latent structure and learn better coordination policies. Our transformer agents perform a graph reasoning over the state of the observable entities. Our transformer Q-mixer learns a monotonic mixing-function from a larger graph that includes the internal and external states of the agents. TransfQMix is designed to be entirely transferable, meaning that same parameters can be used to control and train larger or smaller teams of agents. This enables to deploy promising approaches to save training time and derive general policies in MARL, such as transfer learning, zero-shot transfer, and curriculum learning. We report TransfQMix's performances in the Spread and StarCraft II environments. In both settings, it outperforms state-of-the-art Q-Learning models, and it demonstrates effectiveness in solving problems that other methods can not solve.
Constraint programming is known for being an efficient approach for solving combinatorial problems. Important design choices in a solver are the branching heuristics, which are designed to lead the search to the best solutions in a minimum amount of time. However, developing these heuristics is a time-consuming process that requires problem-specific expertise. This observation has motivated many efforts to use machine learning to automatically learn efficient heuristics without expert intervention. To the best of our knowledge, it is still an open research question. Although several generic variable-selection heuristics are available in the literature, the options for a generic value-selection heuristic are more scarce. In this paper, we propose to tackle this issue by introducing a generic learning procedure that can be used to obtain a value-selection heuristic inside a constraint programming solver. This has been achieved thanks to the combination of a deep Q-learning algorithm, a tailored reward signal, and a heterogeneous graph neural network architecture. Experiments on graph coloring, maximum independent set, and maximum cut problems show that our framework is able to find better solutions close to optimality without requiring a large amounts of backtracks while being generic.
One of the realistic scenarios is taking a sequence of optimal actions to do a task. Reinforcement learning is the most well-known approach to deal with this kind of task in the machine learning community. Finding a suitable alternative could always be an interesting and out-of-the-box matter. Therefore, in this project, we are looking to investigate the capability of NARS and answer the question of whether NARS has the potential to be a substitute for RL or not. Particularly, we are making a comparison between $Q$-Learning and ONA on some environments developed by an Open AI gym. The source code for the experiments is publicly available in the following link: \url{https://github.com/AliBeikmohammadi/OpenNARS-for-Applications/tree/master/misc/Python}.
With the development of experimental quantum technology, quantum control has attracted increasing attention due to the realization of controllable artificial quantum systems. However, because quantum-mechanical systems are often too difficult to analytically deal with, heuristic strategies and numerical algorithms which search for proper control protocols are adopted, and, deep learning, especially deep reinforcement learning (RL), is a promising generic candidate solution for the control problems. Although there have been a few successful applications of deep RL to quantum control problems, most of the existing RL algorithms suffer from instabilities and unsatisfactory reproducibility, and require a large amount of fine-tuning and a large computational budget, both of which limit their applicability. To resolve the issue of instabilities, in this dissertation, we investigate the non-convergence issue of Q-learning. Then, we investigate the weakness of existing convergent approaches that have been proposed, and we develop a new convergent Q-learning algorithm, which we call the convergent deep Q network (C-DQN) algorithm, as an alternative to the conventional deep Q network (DQN) algorithm. We prove the convergence of C-DQN and apply it to the Atari 2600 benchmark. We show that when DQN fail, C-DQN still learns successfully. Then, we apply the algorithm to the measurement-feedback cooling problems of a quantum quartic oscillator and a trapped quantum rigid body. We establish the physical models and analyse their properties, and we show that although both C-DQN and DQN can learn to cool the systems, C-DQN tends to behave more stably, and when DQN suffers from instabilities, C-DQN can achieve a better performance. As the performance of DQN can have a large variance and lack consistency, C-DQN can be a better choice for researches on complicated control problems.
Formulating expert policies as macro actions promises to alleviate the long-horizon issue via structured exploration and efficient credit assignment. However, traditional option-based multi-policy transfer methods suffer from inefficient exploration of macro action's length and insufficient exploitation of useful long-duration macro actions. In this paper, a novel algorithm named EASpace (Enhanced Action Space) is proposed, which formulates macro actions in an alternative form to accelerate the learning process using multiple available sub-optimal expert policies. Specifically, EASpace formulates each expert policy into multiple macro actions with different execution {times}. All the macro actions are then integrated into the primitive action space directly. An intrinsic reward, which is proportional to the execution time of macro actions, is introduced to encourage the exploitation of useful macro actions. The corresponding learning rule that is similar to Intra-option Q-learning is employed to improve the data efficiency. Theoretical analysis is presented to show the convergence of the proposed learning rule. The efficiency of EASpace is illustrated by a grid-based game and a multi-agent pursuit problem. The proposed algorithm is also implemented in physical systems to validate its effectiveness.
We study fair multi-objective reinforcement learning in which an agent must learn a policy that simultaneously achieves high reward on multiple dimensions of a vector-valued reward. Motivated by the fair resource allocation literature, we model this as an expected welfare maximization problem, for some nonlinear fair welfare function of the vector of long-term cumulative rewards. One canonical example of such a function is the Nash Social Welfare, or geometric mean, the log transform of which is also known as the Proportional Fairness objective. We show that even approximately optimal optimization of the expected Nash Social Welfare is computationally intractable even in the tabular case. Nevertheless, we provide a novel adaptation of Q-learning that combines nonlinear scalarized learning updates and non-stationary action selection to learn effective policies for optimizing nonlinear welfare functions. We show that our algorithm is provably convergent, and we demonstrate experimentally that our approach outperforms techniques based on linear scalarization, mixtures of optimal linear scalarizations, or stationary action selection for the Nash Social Welfare Objective.
Safety is still one of the major research challenges in reinforcement learning (RL). In this paper, we address the problem of how to avoid safety violations of RL agents during exploration in probabilistic and partially unknown environments. Our approach combines automata learning for Markov Decision Processes (MDPs) and shield synthesis in an iterative approach. Initially, the MDP representing the environment is unknown. The agent starts exploring the environment and collects traces. From the collected traces, we passively learn MDPs that abstractly represent the safety-relevant aspects of the environment. Given a learned MDP and a safety specification, we construct a shield. For each state-action pair within a learned MDP, the shield computes exact probabilities on how likely it is that executing the action results in violating the specification from the current state within the next $k$ steps. After the shield is constructed, the shield is used during runtime and blocks any actions that induce a too large risk from the agent. The shielded agent continues to explore the environment and collects new data on the environment. Iteratively, we use the collected data to learn new MDPs with higher accuracy, resulting in turn in shields able to prevent more safety violations. We implemented our approach and present a detailed case study of a Q-learning agent exploring slippery Gridworlds. In our experiments, we show that as the agent explores more and more of the environment during training, the improved learned models lead to shields that are able to prevent many safety violations.
Inspired by the cognitive science theory of the explicit human memory systems, we have modeled an agent with short-term, episodic, and semantic memory systems, each of which is modeled with a knowledge graph. To evaluate this system and analyze the behavior of this agent, we designed and released our own reinforcement learning agent environment, "the Room", where an agent has to learn how to encode, store, and retrieve memories to maximize its return by answering questions. We show that our deep Q-learning based agent successfully learns whether a short-term memory should be forgotten, or rather be stored in the episodic or semantic memory systems. Our experiments indicate that an agent with human-like memory systems can outperform an agent without this memory structure in the environment.
Multi-agent reinforcement learning (MARL) suffers from the non-stationarity problem, which is the ever-changing targets at every iteration when multiple agents update their policies at the same time. Starting from first principle, in this paper, we manage to solve the non-stationarity problem by proposing bidirectional action-dependent Q-learning (ACE). Central to the development of ACE is the sequential decision-making process wherein only one agent is allowed to take action at one time. Within this process, each agent maximizes its value function given the actions taken by the preceding agents at the inference stage. In the learning phase, each agent minimizes the TD error that is dependent on how the subsequent agents have reacted to their chosen action. Given the design of bidirectional dependency, ACE effectively turns a multiagent MDP into a single-agent MDP. We implement the ACE framework by identifying the proper network representation to formulate the action dependency, so that the sequential decision process is computed implicitly in one forward pass. To validate ACE, we compare it with strong baselines on two MARL benchmarks. Empirical experiments demonstrate that ACE outperforms the state-of-the-art algorithms on Google Research Football and StarCraft Multi-Agent Challenge by a large margin. In particular, on SMAC tasks, ACE achieves 100% success rate on almost all the hard and super-hard maps. We further study extensive research problems regarding ACE, including extension, generalization, and practicability. Code is made available to facilitate further research.
Increasing traffic demands, higher levels of automation, and communication enhancements provide novel design opportunities for future air traffic controllers (ATCs). This article presents a novel deep reinforcement learning (DRL) controller to aid conflict resolution for autonomous free flight. Although DRL has achieved important advancements in this field, the existing works pay little attention to the explainability and safety issues related to DRL controllers, particularly the safety under adversarial attacks. To address those two issues, we design a fully explainable DRL framework wherein we: 1) decompose the coupled Q value learning model into a safety-awareness and efficiency (reach the target) one; and 2) use information from surrounding intruders as inputs, eliminating the needs of central controllers. In our simulated experiments, we show that by decoupling the safety-awareness and efficiency, we can exceed performance on free flight control tasks while dramatically improving explainability on practical. In addition, the safety Q learning module provides rich information about the safety situation of environments. To study the safety under adversarial attacks, we additionally propose an adversarial attack strategy that can impose both safety-oriented and efficiency-oriented attacks. The adversarial aims to minimize safety/efficiency by only attacking the agent at a few time steps. In the experiments, our attack strategy increases as many collisions as the uniform attack (i.e., attacking at every time step) by only attacking the agent four times less often, which provide insights into the capabilities and restrictions of the DRL in future ATC designs. The source code is publicly available at https://github.com/WLeiiiii/Gym-ATC-Attack-Project.
This effort is focused on examining the behavior of reinforcement learning systems in personalization environments and detailing the differences in policy entropy associated with the type of learning algorithm utilized. We demonstrate that Policy Optimization agents often possess low-entropy policies during training, which in practice results in agents prioritizing certain actions and avoiding others. Conversely, we also show that Q-Learning agents are far less susceptible to such behavior and generally maintain high-entropy policies throughout training, which is often preferable in real-world applications. We provide a wide range of numerical experiments as well as theoretical justification to show that these differences in entropy are due to the type of learning being employed.
Reinforcement learning (RL) is gaining popularity as an effective approach for traffic signal control (TSC) and is increasingly applied in this domain. However, most existing RL methodologies are confined to a single-stage TSC framework, primarily focusing on selecting an appropriate traffic signal phase at fixed action intervals, leading to inflexible and less adaptable phase durations. To address such limitations, we introduce a novel two-stage TSC framework named DynamicLight. This framework initiates with a phase control strategy responsible for determining the optimal traffic phase, followed by a duration control strategy tasked with determining the corresponding phase duration. Experimental results show that DynamicLight outperforms state-of-the-art TSC models and exhibits exceptional model generalization capabilities. Additionally, the robustness and potential for real-world implementation of DynamicLight are further demonstrated and validated through various DynamicLight variants. The code is released at https://github.com/LiangZhang1996/DynamicLight.
While there has been substantial success for solving continuous control with actor-critic methods, simpler critic-only methods such as Q-learning find limited application in the associated high-dimensional action spaces. However, most actor-critic methods come at the cost of added complexity: heuristics for stabilisation, compute requirements and wider hyperparameter search spaces. We show that a simple modification of deep Q-learning largely alleviates these issues. By combining bang-bang action discretization with value decomposition, framing single-agent control as cooperative multi-agent reinforcement learning (MARL), this simple critic-only approach matches performance of state-of-the-art continuous actor-critic methods when learning from features or pixels. We extend classical bandit examples from cooperative MARL to provide intuition for how decoupled critics leverage state information to coordinate joint optimization, and demonstrate surprisingly strong performance across a variety of continuous control tasks.
We consider a hybrid reinforcement learning setting (Hybrid RL), in which an agent has access to an offline dataset and the ability to collect experience via real-world online interaction. The framework mitigates the challenges that arise in both pure offline and online RL settings, allowing for the design of simple and highly effective algorithms, in both theory and practice. We demonstrate these advantages by adapting the classical Q learning/iteration algorithm to the hybrid setting, which we call Hybrid Q-Learning or Hy-Q. In our theoretical results, we prove that the algorithm is both computationally and statistically efficient whenever the offline dataset supports a high-quality policy and the environment has bounded bilinear rank. Notably, we require no assumptions on the coverage provided by the initial distribution, in contrast with guarantees for policy gradient/iteration methods. In our experimental results, we show that Hy-Q with neural network function approximation outperforms state-of-the-art online, offline, and hybrid RL baselines on challenging benchmarks, including Montezuma's Revenge.
Recently, auto-bidding technique has become an essential tool to increase the revenue of advertisers. Facing the complex and ever-changing bidding environments in the real-world advertising system (RAS), state-of-the-art auto-bidding policies usually leverage reinforcement learning (RL) algorithms to generate real-time bids on behalf of the advertisers. Due to safety concerns, it was believed that the RL training process can only be carried out in an offline virtual advertising system (VAS) that is built based on the historical data generated in the RAS. In this paper, we argue that there exists significant gaps between the VAS and RAS, making the RL training process suffer from the problem of inconsistency between online and offline (IBOO). Firstly, we formally define the IBOO and systematically analyze its causes and influences. Then, to avoid the IBOO, we propose a sustainable online RL (SORL) framework that trains the auto-bidding policy by directly interacting with the RAS, instead of learning in the VAS. Specifically, based on our proof of the Lipschitz smooth property of the Q function, we design a safe and efficient online exploration (SER) policy for continuously collecting data from the RAS. Meanwhile, we derive the theoretical lower bound on the safety of the SER policy. We also develop a variance-suppressed conservative Q-learning (V-CQL) method to effectively and stably learn the auto-bidding policy with the collected data. Finally, extensive simulated and real-world experiments validate the superiority of our approach over the state-of-the-art auto-bidding algorithm.
Progress in deep learning highlights the tremendous potential of utilizing diverse robotic datasets for attaining effective generalization and makes it enticing to consider leveraging broad datasets for attaining robust generalization in robotic learning as well. However, in practice, we often want to learn a new skill in a new environment that is unlikely to be contained in the prior data. Therefore we ask: how can we leverage existing diverse offline datasets in combination with small amounts of task-specific data to solve new tasks, while still enjoying the generalization benefits of training on large amounts of data? In this paper, we demonstrate that end-to-end offline RL can be an effective approach for doing this, without the need for any representation learning or vision-based pre-training. We present pre-training for robots (PTR), a framework based on offline RL that attempts to effectively learn new tasks by combining pre-training on existing robotic datasets with rapid fine-tuning on a new task, with as few as 10 demonstrations. PTR utilizes an existing offline RL method, conservative Q-learning (CQL), but extends it to include several crucial design decisions that enable PTR to actually work and outperform a variety of prior methods. To our knowledge, PTR is the first RL method that succeeds at learning new tasks in a new domain on a real WidowX robot with as few as 10 task demonstrations, by effectively leveraging an existing dataset of diverse multi-task robot data collected in a variety of toy kitchens. We also demonstrate that PTR can enable effective autonomous fine-tuning and improvement in a handful of trials, without needing any demonstrations. An accompanying overview video can be found in the supplementary material and at thi URL: https://sites.google.com/view/ptr-final/