Effectively leveraging large, previously collected datasets in reinforcement learning (RL) is a key challenge for large-scale real-world applications. Offline RL algorithms promise to learn effective policies from previously-collected, static datasets without further interaction. However, in practice, offline RL presents a major challenge, and standard off-policy RL methods can fail due to overestimation of values induced by the distributional shift between the dataset and the learned policy, especially when training on complex and multi-modal data distributions. In this paper, we propose conservative Q-learning (CQL), which aims to address these limitations by learning a conservative Q-function such that the expected value of a policy under this Q-function lower-bounds its true value. We theoretically show that CQL produces a lower bound on the value of the current policy and that it can be incorporated into a policy learning procedure with theoretical improvement guarantees. In practice, CQL augments the standard Bellman error objective with a simple Q-value regularizer which is straightforward to implement on top of existing deep Q-learning and actor-critic implementations. On both discrete and continuous control domains, we show that CQL substantially outperforms existing offline RL methods, often learning policies that attain 2-5 times higher final return, especially when learning from complex and multi-modal data distributions.
Reinforcement learning is a popular machine learning paradigm which can find near optimal solutions to complex problems. Most often, these procedures involve function approximation using neural networks with gradient based updates to optimise weights for the problem being considered. While this common approach generally works well, there are other update mechanisms which are largely unexplored in reinforcement learning. One such mechanism is Extreme Learning Machines. These were initially proposed to drastically improve the training speed of neural networks and have since seen many applications. Here we attempt to apply extreme learning machines to a reinforcement learning problem in the same manner as gradient based updates. This new algorithm is called Extreme Q-Learning Machine (EQLM). We compare its performance to a typical Q-Network on the cart-pole task - a benchmark reinforcement learning problem - and show EQLM has similar long-term learning performance to a Q-Network.
Centralized training with decentralized execution has become an important paradigm in multi-agent learning. Though practical, current methods rely on restrictive assumptions to decompose the centralized value function across agents for execution. In this paper, we eliminate this restriction by proposing multi-agent determinantal Q-learning. Our method is established on Q-DPP, an extension of determinantal point process (DPP) with partition-matroid constraint to multi-agent setting. Q-DPP promotes agents to acquire diverse behavioral models; this allows a natural factorization of the joint Q-functions with no need for \emph{a priori} structural constraints on the value function or special network architectures. We demonstrate that Q-DPP generalizes major solutions including VDN, QMIX, and QTRAN on decentralizable cooperative tasks. To efficiently draw samples from Q-DPP, we adopt an existing sample-by-projection sampler with theoretical approximation guarantee. The sampler also benefits exploration by coordinating agents to cover orthogonal directions in the state space during multi-agent training. We evaluate our algorithm on various cooperative benchmarks; its effectiveness has been demonstrated when compared with the state-of-the-art.
Penetration testing is a security exercise aimed at assessing the security of a system by simulating attacks against it. So far, penetration testing has been carried out mainly by trained human attackers and its success critically depended on the available expertise. Automating this practice constitutes a non-trivial problem, as the range of actions that a human expert may attempts against a system and the range of knowledge she relies on to take her decisions are hard to capture. In this paper, we focus our attention on simplified penetration testing problems expressed in the form of capture the flag hacking challenges, and we analyze how model-free reinforcement learning algorithms may help to solve them. In modeling these capture the flag competitions as reinforcement learning problems we highlight that a specific challenge that characterize penetration testing is the problem of discovering the structure of the problem at hand. We then show how this challenge may be eased by relying on different forms of prior knowledge that may be provided to the agent. In this way we demonstrate how the feasibility of tackling penetration testing using reinforcement learning may rest on a careful trade-off between model-free and model-based algorithms. By using techniques to inject a priori knowledge, we show it is possible to better direct the agent and restrict the space of its exploration problem, thus achieving solutions more efficiently.
Typical end-to-end formulations for learning robotic navigation involve predicting a small set of steering command actions (e.g., step forward, turn left, turn right, etc.) from images of the current state (e.g., a bird's-eye view of a SLAM reconstruction). Instead, we show that it can be advantageous to learn with dense action representations defined in the same domain as the state. In this work, we present "spatial action maps," in which the set of possible actions is represented by a pixel map (aligned with the input image of the current state), where each pixel represents a local navigational endpoint at the corresponding scene location. Using ConvNets to infer spatial action maps from state images, action predictions are thereby spatially anchored on local visual features in the scene, enabling significantly faster learning of complex behaviors for mobile manipulation tasks with reinforcement learning. In our experiments, we task a robot with pushing objects to a goal location, and find that policies learned with spatial action maps achieve much better performance than traditional alternatives.
Reinforcement learning agents learn by encouraging behaviours which maximize their total reward, usually provided by the environment. In many environments, however, the reward is provided after a series of actions rather than each single action, leading the agent to experience ambiguity in terms of whether those actions are effective, an issue known as the credit assignment problem. In this paper, we propose two strategies inspired by behavioural psychology to enable the agent to intrinsically estimate more informative reward values for actions with no reward. The first strategy, called self-punishment (SP), discourages the agent from making mistakes that lead to undesirable terminal states. The second strategy, called the rewards backfill (RB), backpropagates the rewards between two rewarded actions. We prove that, under certain assumptions and regardless of the reinforcement learning algorithm used, these two strategies maintain the order of policies in the space of all possible policies in terms of their total reward, and, by extension, maintain the optimal policy. Hence, our proposed strategies integrate with any reinforcement learning algorithm that learns a value or action-value function through experience. We incorporated these two strategies into three popular deep reinforcement learning approaches and evaluated the results on thirty Atari games. After parameter tuning, our results indicate that the proposed strategies improve the tested methods in over 65 percent of tested games by up to over 25 times performance improvement.
The study of unsupervised learning can be generally divided into two categories: imitation learning and reinforcement learning. In imitation learning the machine learns by mimicking the behavior of an expert system whereas in reinforcement learning the machine learns via direct environment feedback. Traditional deep reinforcement learning takes a significant time before the machine starts to converge to an optimal policy. This paper proposes Augmented Q-Imitation-Learning, a method by which deep reinforcement learning convergence can be accelerated by applying Q-imitation-learning as the initial training process in traditional Deep Q-learning.
Reinforcement learning is one of the most popular approaches for automated game playing. This method allows an agent to estimate the expected utility of its state in order to make optimal actions in an unknown environment. We seek to apply reinforcement learning algorithms to the game Flappy Bird. We implement SARSA and Q-Learning with some modifications such as $\epsilon$-greedy policy, discretization and backward updates. We find that SARSA and Q-Learning outperform the baseline, regularly achieving scores of 1400+, with the highest in-game score of 2069.
We propose FACtored Multi-Agent Centralised policy gradients (FACMAC), a new method for cooperative multi-agent reinforcement learning in both discrete and continuous action spaces. Like MADDPG, a popular multi-agent actor-critic method, our approach uses deep deterministic policy gradients to learn policies. However, FACMAC learns a centralised but factored critic, which combines per-agent utilities into the joint action-value function via a non-linear monotonic function, as in QMIX, a popular multi-agent Q-learning algorithm. However, unlike QMIX, there are no inherent constraints on factoring the critic. We thus also employ a nonmonotonic factorisation and empirically demonstrate that its increased representational capacity allows it to solve some tasks that cannot be solved with monolithic, or monotonically factored critics. In addition, FACMAC uses a centralised policy gradient estimator that optimises over the entire joint action space, rather than optimising over each agent's action space separately as in MADDPG. This allows for more coordinated policy changes and fully reaps the benefits of a centralised critic. We evaluate FACMAC on variants of the multi-agent particle environments, a novel multi-agent MuJoCo benchmark, and a challenging set of StarCraft II micromanagement tasks. Empirical results demonstrate FACMAC's superior performance over MADDPG and other baselines on all three domains.
Optimistic initialisation is an effective strategy for efficient exploration in reinforcement learning (RL). In the tabular case, all provably efficient model-free algorithms rely on it. However, model-free deep RL algorithms do not use optimistic initialisation despite taking inspiration from these provably efficient tabular algorithms. In particular, in scenarios with only positive rewards, Q-values are initialised at their lowest possible values due to commonly used network initialisation schemes, a pessimistic initialisation. Merely initialising the network to output optimistic Q-values is not enough, since we cannot ensure that they remain optimistic for novel state-action pairs, which is crucial for exploration. We propose a simple count-based augmentation to pessimistically initialised Q-values that separates the source of optimism from the neural network. We show that this scheme is provably efficient in the tabular setting and extend it to the deep RL setting. Our algorithm, Optimistic Pessimistically Initialised Q-Learning (OPIQ), augments the Q-value estimates of a DQN-based agent with count-derived bonuses to ensure optimism during both action selection and bootstrapping. We show that OPIQ outperforms non-optimistic DQN variants that utilise a pseudocount-based intrinsic motivation in hard exploration tasks, and that it predicts optimistic estimates for novel state-action pairs.
Q-learning suffers from overestimation bias, because it approximates the maximum action value using the maximum estimated action value. Algorithms have been proposed to reduce overestimation bias, but we lack an understanding of how bias interacts with performance, and the extent to which existing algorithms mitigate bias. In this paper, we 1) highlight that the effect of overestimation bias on learning efficiency is environment-dependent; 2) propose a generalization of Q-learning, called \emph{Maxmin Q-learning}, which provides a parameter to flexibly control bias; 3) show theoretically that there exists a parameter choice for Maxmin Q-learning that leads to unbiased estimation with a lower approximation variance than Q-learning; and 4) prove the convergence of our algorithm in the tabular case, as well as convergence of several previous Q-learning variants, using a novel Generalized Q-learning framework. We empirically verify that our algorithm better controls estimation bias in toy environments, and that it achieves superior performance on several benchmark problems.
Can simple algorithms with a good representation solve challenging reinforcement learning problems? In this work, we answer this question in the affirmative, where we take "simple learning algorithm" to be tabular Q-Learning, the "good representations" to be a learned state abstraction, and "challenging problems" to be continuous control tasks. Our main contribution is a learning algorithm that abstracts a continuous state-space into a discrete one. We transfer this learned representation to unseen problems to enable effective learning. We provide theory showing that learned abstractions maintain a bounded value loss, and we report experiments showing that the abstractions empower tabular Q-Learning to learn efficiently in unseen tasks.
We consider the problem of energy management in microgrid networks. A microgrid is capable of generating a limited amount of energy from a renewable resource and is responsible for handling the demands of its dedicated customers. Owing to the variable nature of renewable generation and the demands of the customers, it becomes imperative that each microgrid optimally manages its energy. This involves intelligently scheduling the demands at the customer side, selling (when there is a surplus) and buying (when there is a deficit) the power from its neighboring microgrids depending on its current and future needs. Typically, the transaction of power among the microgrids happens at a pre-decided price by the central grid. In this work, we formulate the problems of demand and battery scheduling, energy trading and dynamic pricing (where we allow the microgrids to decide the price of the transaction depending on their current configuration of demand and renewable energy) in the framework of stochastic games. Subsequently, we propose a novel approach that makes use of independent learners Deep Q-learning algorithm to solve this problem. Through extensive empirical evaluation, we show that our proposed framework is more beneficial to the majority of the microgrids and we provide a detailed analysis of the results.
It is difficult to be able to imitate well in unknown states from a small amount of expert data and sampling data. Supervised learning methods such as Behavioral Cloning do not require sampling data, but usually suffer from distribution shift. The methods based on reinforcement learning, such as inverse reinforcement learning and generative adversarial imitation learning (GAIL), can learn from only a few expert data. However, they often need to interact with the environment. Soft Q imitation learning addressed the problems, and it was shown that it could learn efficiently by combining Behavioral Cloning and soft Q-learning with constant rewards. In order to make this algorithm more robust to distribution shift, we propose Discriminator Soft Actor Critic (DSAC). It uses a reward function based on adversarial inverse reinforcement learning instead of constant rewards. We evaluated it on PyBullet environments with only four expert trajectories.
Signalized intersections are managed by controllers that assign right of way (green, yellow, and red lights) to non-conflicting directions. Optimizing the actuation policy of such controllers is expected to alleviate traffic congestion and its adverse impact. Given such a safety-critical domain, the affiliated actuation policy is required to be interpretable in a way that can be understood and regulated by a human. This paper presents and analyzes several on-line optimization techniques for tuning interpretable control functions. Although these techniques are defined in a general way, this paper assumes a specific class of interpretable control functions (polynomial functions) for analysis purposes. We show that such an interpretable policy function can be as effective as a deep neural network for approximating an optimized signal actuation policy. We present empirical evidence that supports the use of value-based reinforcement learning for on-line training of the control function. Specifically, we present and study three variants of the Deep Q-learning algorithm that allow the training of an interpretable policy function. Our Deep Regulatable Hardmax Q-learning variant is shown to be particularly effective in optimizing our interpretable actuation policy, resulting in up to 19.4% reduced vehicles delay compared to commonly deployed actuated signal controllers.
Human players in professional team sports achieve high level coordination by dynamically choosing complementary skills and executing primitive actions to perform these skills. As a step toward creating intelligent agents with this capability for fully cooperative multi-agent settings, we propose a two-level hierarchical multi-agent reinforcement learning (MARL) algorithm with unsupervised skill discovery. Agents learn useful and distinct skills at the low level via independent Q-learning, while they learn to select complementary latent skill variables at the high level via centralized multi-agent training with an extrinsic team reward. The set of low-level skills emerges from an intrinsic reward that solely promotes the decodability of latent skill variables from the trajectory of a low-level skill, without the need for hand-crafted rewards for each skill. For scalable decentralized execution, each agent independently chooses latent skill variables and primitive actions based on local observations. Our overall method enables the use of general cooperative MARL algorithms for training high level policies and single-agent RL for training low level skills. Experiments on a stochastic high dimensional team game show the emergence of useful skills and cooperative team play. The interpretability of the learned skills show the promise of the proposed method for achieving human-AI cooperation in team sports games.
Join query optimization is a complex task and is central to the performance of query processing. In fact it belongs to the class of NP-hard problems. Traditional query optimizers use dynamic programming (DP) methods combined with a set of rules and restrictions to avoid exhaustive enumeration of all possible join orders. However, DP methods are very resource intensive. Moreover, given simplifying assumptions of attribute independence, traditional query optimizers rely on erroneous cost estimations, which can lead to suboptimal query plans. Recent success of deep reinforcement learning (DRL) creates new opportunities for the field of query optimization to tackle the above-mentioned problems. In this paper, we present our DRL-based Fully Observed Optimizer (FOOP) which is a generic query optimization framework that enables plugging in different machine learning algorithms. The main idea of FOOP is to use a data-adaptive learning query optimizer that avoids exhaustive enumerations of join orders and is thus significantly faster than traditional approaches based on dynamic programming. In particular, we evaluate various DRL-algorithms and show that Proximal Policy Optimization significantly outperforms Q-learning based algorithms. Finally we demonstrate how ensemble learning techniques combined with DRL can further improve the query optimizer.
Unmanned Aerial Vehicles (UAVs), autonomously-guided aircraft, are widely used for tasks involving surveillance and reconnaissance. A version of the pursuit-evasion problems centered around UAVs and its variants has been extensively studied in recent years due to numerous breakthroughs in AI. We present an approach to UAV pursuit-evasion in a 2D aerial-engagement environment using reinforcement learning (RL), a machine learning paradigm concerned with goal-oriented algorithms. In this work, a UAV wielding the greedy shooter strategy engages with a UAV trained using deep Q-learning techniques. Simulated results show that the latter UAV wins every engagement in which the UAVs are suffciently separated during initialization. This approach highlights an exhaustive and robust application of reinforcement learning to pursuit-evasion that provides insight into effective strategies for UAV flight and interaction.
In this paper, we derive a generalization of the Speedy Q-learning (SQL) algorithm that was proposed in the Reinforcement Learning (RL) literature to handle slow convergence of Watkins' Q-learning. In most RL algorithms such as Q-learning, the Bellman equation and the Bellman operator play an important role. It is possible to generalize the Bellman operator using the technique of successive relaxation. We use the generalized Bellman operator to derive a simple and efficient family of algorithms called Generalized Speedy Q-learning (GSQL-w) and analyze its finite time performance. We show that GSQL-w has an improved finite time performance bound compared to SQL for the case when the relaxation parameter w is greater than 1. This improvement is a consequence of the contraction factor of the generalized Bellman operator being less than that of the standard Bellman operator. Numerical experiments are provided to demonstrate the empirical performance of the GSQL-w algorithm.
There has recently been a surge in research in batch Deep Reinforcement Learning (DRL), which aims for learning a high-performing policy from a given dataset without additional interactions with the environment. We propose a new algorithm, Best-Action Imitation Learning (BAIL), which strives for both simplicity and performance. BAIL learns a V function, uses the V function to select actions it believes to be high-performing, and then uses those actions to train a policy network using imitation learning. For the MuJoCo benchmark, we provide a comprehensive experimental study of BAIL, comparing its performance to four other batch Q-learning and imitation-learning schemes for a large variety of batch datasets. Our experiments show that BAIL's performance is much higher than the other schemes, and is also computationally much faster than the batch Q-learning schemes.