In this paper, we consider cooperative multi-agent reinforcement learning (MARL) with sparse reward. To tackle this problem, we propose a novel method named MASER: MARL with subgoals generated from experience replay buffer. Under the widely-used assumption of centralized training with decentralized execution and consistent Q-value decomposition for MARL, MASER automatically generates proper subgoals for multiple agents from the experience replay buffer by considering both individual Q-value and total Q-value. Then, MASER designs individual intrinsic reward for each agent based on actionable representation relevant to Q-learning so that the agents reach their subgoals while maximizing the joint action value. Numerical results show that MASER significantly outperforms StarCraft II micromanagement benchmark compared to other state-of-the-art MARL algorithms.
Context: Machine learning (ML) may enable effective automated test generation. Objective: We characterize emerging research, examining testing practices, researcher goals, ML techniques applied, evaluation, and challenges. Methods: We perform a systematic mapping on a sample of 124 publications. Results: ML generates input for system, GUI, unit, performance, and combinatorial testing or improves the performance of existing generation methods. ML is also used to generate test verdicts, property-based, and expected output oracles. Supervised learning - often based on neural networks - and reinforcement learning - often based on Q-learning - are common, and some publications also employ unsupervised or semi-supervised learning. (Semi-/Un-)Supervised approaches are evaluated using both traditional testing metrics and ML-related metrics (e.g., accuracy), while reinforcement learning is often evaluated using testing metrics tied to the reward function. Conclusion: Work-to-date shows great promise, but there are open challenges regarding training data, retraining, scalability, evaluation complexity, ML algorithms employed - and how they are applied - benchmarks, and replicability. Our findings can serve as a roadmap and inspiration for researchers in this field.
Since reinforcement learning algorithms are notoriously data-intensive, the task of sampling observations from the environment is usually split across multiple agents. However, transferring these observations from the agents to a central location can be prohibitively expensive in terms of communication cost, and it can also compromise the privacy of each agent's local behavior policy. Federated reinforcement learning is a framework in which $N$ agents collaboratively learn a global model, without sharing their individual data and policies. This global model is the unique fixed point of the average of $N$ local operators, corresponding to the $N$ agents. Each agent maintains a local copy of the global model and updates it using locally sampled data. In this paper, we show that by careful collaboration of the agents in solving this joint fixed point problem, we can find the global model $N$ times faster, also known as linear speedup. We first propose a general framework for federated stochastic approximation with Markovian noise and heterogeneity, showing linear speedup in convergence. We then apply this framework to federated reinforcement learning algorithms, examining the convergence of federated on-policy TD, off-policy TD, and $Q$-learning.
Massive practical works addressed by Deep Q-network (DQN) algorithm have indicated that stochastic policy, despite its simplicity, is the most frequently used exploration approach. However, most existing stochastic exploration approaches either explore new actions heuristically regardless of Q-values or inevitably introduce bias into the learning process to couple the sampling with Q-values. In this paper, we propose a novel preference-guided $\epsilon$-greedy exploration algorithm that can efficiently learn the action distribution in line with the landscape of Q-values for DQN without introducing additional bias. Specifically, we design a dual architecture consisting of two branches, one of which is a copy of DQN, namely the Q-branch. The other branch, which we call the preference branch, learns the action preference that the DQN implicit follows. We theoretically prove that the policy improvement theorem holds for the preference-guided $\epsilon$-greedy policy and experimentally show that the inferred action preference distribution aligns with the landscape of corresponding Q-values. Consequently, preference-guided $\epsilon$-greedy exploration motivates the DQN agent to take diverse actions, i.e., actions with larger Q-values can be sampled more frequently whereas actions with smaller Q-values still have a chance to be explored, thus encouraging the exploration. We assess the proposed method with four well-known DQN variants in nine different environments. Extensive results confirm the superiority of our proposed method in terms of performance and convergence speed. Index Terms- Preference-guided exploration, stochastic policy, data efficiency, deep reinforcement learning, deep Q-learning.
Deep Reinforcement Learning (DRL) algorithms have been increasingly employed during the last decade to solve various decision-making problems such as autonomous driving and robotics. However, these algorithms have faced great challenges when deployed in safety-critical environments since they often exhibit erroneous behaviors that can lead to potentially critical errors. One way to assess the safety of DRL agents is to test them to detect possible faults leading to critical failures during their execution. This raises the question of how we can efficiently test DRL policies to ensure their correctness and adherence to safety requirements. Most existing works on testing DRL agents use adversarial attacks that perturb states or actions of the agent. However, such attacks often lead to unrealistic states of the environment. Their main goal is to test the robustness of DRL agents rather than testing the compliance of agents' policies with respect to requirements. Due to the huge state space of DRL environments, the high cost of test execution, and the black-box nature of DRL algorithms, the exhaustive testing of DRL agents is impossible. In this paper, we propose a Search-based Testing Approach of Reinforcement Learning Agents (STARLA) to test the policy of a DRL agent by effectively searching for failing executions of the agent within a limited testing budget. We use machine learning models and a dedicated genetic algorithm to narrow the search towards faulty episodes. We apply STARLA on Deep-Q-Learning agents which are widely used as benchmarks and show that it significantly outperforms Random Testing by detecting more faults related to the agent's policy. We also investigate how to extract rules that characterize faulty episodes of the DRL agent using our search results. Such rules can be used to understand the conditions under which the agent fails and thus assess its deployment risks.
Test point insertion (TPI) is a widely used technique for testability enhancement, especially for logic built-in self-test (LBIST) due to its relatively low fault coverage. In this paper, we propose a novel TPI approach based on deep reinforcement learning (DRL), named DeepTPI. Unlike previous learning-based solutions that formulate the TPI task as a supervised-learning problem, we train a novel DRL agent, instantiated as the combination of a graph neural network (GNN) and a Deep Q-Learning network (DQN), to maximize the test coverage improvement. Specifically, we model circuits as directed graphs and design a graph-based value network to estimate the action values for inserting different test points. The policy of the DRL agent is defined as selecting the action with the maximum value. Moreover, we apply the general node embeddings from a pre-trained model to enhance node features, and propose a dedicated testability-aware attention mechanism for the value network. Experimental results on circuits with various scales show that DeepTPI significantly improves test coverage compared to the commercial DFT tool. The code of this work is available at https://github.com/cure-lab/DeepTPI.
While reinforcement learning (RL) from raw images has been largely investigated in the last decade, existing approaches still suffer from a number of constraints. The high input dimension is often handled using either expert knowledge to extract handcrafted features or environment encoding through convolutional networks. Both solutions require numerous parameters to be optimized. In contrast, we propose a generic method to extract sparse features from raw images with few trainable parameters. We achieved this using a Radial Basis Function Network (RBFN) directly on raw image. We evaluate the performance of the proposed approach for visual extraction in Q-learning tasks in the Vizdoom environment. Then, we compare our results with two Deep Q-Network, one trained directly on images and another one trained on feature extracted by a pretrained auto-encoder. We show that the proposed approach provides similar or, in some cases, even better performances with fewer trainable parameters while being conceptually simpler.
The study of electromagnetic detection satellite scheduling problem (EDSSP) has attracted attention due to the detection requirements for a large number of targets. This paper proposes a mixed-integer programming model for the EDSSP problem and a genetic algorithm based on reinforcement learning (RL-GA). Numerous factors that affect electromagnetic detection are considered in the model, such as detection mode, bandwidth, and other factors. The RL-GA embeds a Q-learning method into an improved genetic algorithm, and the evolution of each individual depends on the decision of the agent. Q-learning is used to guide the population search process by choosing evolution operators. In this way, the search information can be effectively used by the reinforcement learning method. In the algorithm, we design a reward function to update the Q value. According to the problem characteristics, a new combination of <state, action> is proposed. The RL-GA also uses an elite individual retention strategy to improve search performance. After that, a task time window selection algorithm (TTWSA) is proposed to evaluate the performance of population evolution. Several experiments are used to examine the scheduling effect of the proposed algorithm. Through the experimental verification of multiple instances, it can be seen that the RL-GA can solve the EDSSP problem effectively. Compared with the state-of-the-art algorithms, the RL-GA performs better in several aspects.
With the digitalization of the financial market, dealers are increasingly handling market-making activities by algorithms. Recent antitrust literature raises concerns on collusion caused by artificial intelligence. This paper studies the possibility of cooperation between market makers via independent Q-learning. Market making with inventory risk is formulated as a repeated general-sum game. Under a stag-hunt type payoff, we find that market makers can learn cooperative strategies without communication. In general, high spreads can have the largest probability even when the lowest spread is the unique Nash equilibrium. Moreover, introducing more agents into the game does not necessarily eliminate the presence of supra-competitive spreads.
Offline reinforcement learning (RL) defines the task of learning from a static logged dataset without continually interacting with the environment. The distribution shift between the learned policy and the behavior policy makes it necessary for the value function to stay conservative such that out-of-distribution (OOD) actions will not be severely overestimated. However, existing approaches, penalizing the unseen actions or regularizing with the behavior policy, are too pessimistic, which suppresses the generalization of the value function and hinders the performance improvement. This paper explores mild but enough conservatism for offline learning while not harming generalization. We propose Mildly Conservative Q-learning (MCQ), where OOD actions are actively trained by assigning them proper pseudo Q values. We theoretically show that MCQ induces a policy that behaves at least as well as the behavior policy and no erroneous overestimation will occur for OOD actions. Experimental results on the D4RL benchmarks demonstrate that MCQ achieves remarkable performance compared with prior work. Furthermore, MCQ shows superior generalization ability when transferring from offline to online, and significantly outperforms baselines. Our code is publicly available at https://github.com/dmksjfl/MCQ.
Unit commitment (UC) is a fundamental problem in the day-ahead electricity market, and it is critical to solve UC problems efficiently. Mathematical optimization techniques like dynamic programming, Lagrangian relaxation, and mixed-integer quadratic programming (MIQP) are commonly adopted for UC problems. However, the calculation time of these methods increases at an exponential rate with the amount of generators and energy resources, which is still the main bottleneck in industry. Recent advances in artificial intelligence have demonstrated the capability of reinforcement learning (RL) to solve UC problems. Unfortunately, the existing research on solving UC problems with RL suffers from the curse of dimensionality when the size of UC problems grows. To deal with these problems, we propose an optimization method-assisted ensemble deep reinforcement learning algorithm, where UC problems are formulated as a Markov Decision Process (MDP) and solved by multi-step deep Q-learning in an ensemble framework. The proposed algorithm establishes a candidate action set by solving tailored optimization problems to ensure a relatively high performance and the satisfaction of operational constraints. Numerical studies on IEEE 118 and 300-bus systems show that our algorithm outperforms the baseline RL algorithm and MIQP. Furthermore, the proposed algorithm shows strong generalization capacity under unforeseen operational conditions.
We present an empirical study on the use of continual learning (CL) methods in a reinforcement learning (RL) scenario, which, to the best of our knowledge, has not been described before. CL is a very active recent research topic concerned with machine learning under non-stationary data distributions. Although this naturally applies to RL, the use of dedicated CL methods is still uncommon. This may be due to the fact that CL methods often assume a decomposition of CL problems into disjoint sub-tasks of stationary distribution, that the onset of these sub-tasks is known, and that sub-tasks are non-contradictory. In this study, we perform an empirical comparison of selected CL methods in a RL problem where a physically simulated robot must follow a racetrack by vision. In order to make CL methods applicable, we restrict the RL setting and introduce non-conflicting subtasks of known onset, which are however not disjoint and whose distribution, from the learner's point of view, is still non-stationary. Our results show that dedicated CL methods can significantly improve learning when compared to the baseline technique of "experience replay".
We derive a concentration bound for a Q-learning algorithm for average cost Markov decision processes based on an equivalent shortest path problem, and compare it numerically with the alternative scheme based on relative value iteration.
In reinforcement learning (RL), experience replay-based sampling techniques play a crucial role in promoting convergence by eliminating spurious correlations. However, widely used methods such as uniform experience replay (UER) and prioritized experience replay (PER) have been shown to have sub-optimal convergence and high seed sensitivity respectively. To address these issues, we propose a novel approach called IntrospectiveExperience Replay (IER) that selectively samples batches of data points prior to surprising events. Our method builds upon the theoretically sound reverse experience replay (RER) technique, which has been shown to reduce bias in the output of Q-learning-type algorithms with linear function approximation. However, this approach is not always practical or reliable when using neural function approximation. Through empirical evaluations, we demonstrate that IER with neural function approximation yields reliable and superior performance compared toUER, PER, and hindsight experience replay (HER) across most tasks.
The $Q$-learning algorithm is a simple and widely-used stochastic approximation scheme for reinforcement learning, but the basic protocol can exhibit instability in conjunction with function approximation. Such instability can be observed even with linear function approximation. In practice, tools such as target networks and experience replay appear to be essential, but the individual contribution of each of these mechanisms is not well understood theoretically. This work proposes an exploration variant of the basic $Q$-learning protocol with linear function approximation. Our modular analysis illustrates the role played by each algorithmic tool that we adopt: a second order update rule, a set of target networks, and a mechanism akin to experience replay. Together, they enable state of the art regret bounds on linear MDPs while preserving the most prominent feature of the algorithm, namely a space complexity independent of the number of step elapsed. We show that the performance of the algorithm degrades very gracefully under a novel and more permissive notion of approximation error. The algorithm also exhibits a form of instance-dependence, in that its performance depends on the "effective" feature dimension.
Learning the underlying equation from data is a fundamental problem in many disciplines. Recent advances rely on Neural Networks (NNs) but do not provide theoretical guarantees in obtaining the exact equations owing to the non-convexity of NNs. In this paper, we propose Convex Neural Symbolic Learning (CoNSoLe) to seek convexity under mild conditions. The main idea is to decompose the recovering process into two steps and convexify each step. In the first step of searching for right symbols, we convexify the deep Q-learning. The key is to maintain double convexity for both the negative Q-function and the negative reward function in each iteration, leading to provable convexity of the negative optimal Q function to learn the true symbol connections. Conditioned on the exact searching result, we construct a Locally Convex equation Learner (LoCaL) neural network to convexify the estimation of symbol coefficients. With such a design, we quantify a large region with strict convexity in the loss surface of LoCaL for commonly used physical functions. Finally, we demonstrate the superior performance of the CoNSoLe framework over the state-of-the-art on a diverse set of datasets.
The successes of deep Reinforcement Learning (RL) are limited to settings where we have a large stream of online experiences, but applying RL in the data-efficient setting with limited access to online interactions is still challenging. A key to data-efficient RL is good value estimation, but current methods in this space fail to fully utilise the structure of the trajectory data gathered from the environment. In this paper, we treat the transition data of the MDP as a graph, and define a novel backup operator, Graph Backup, which exploits this graph structure for better value estimation. Compared to multi-step backup methods such as $n$-step $Q$-Learning and TD($\lambda$), Graph Backup can perform counterfactual credit assignment and gives stable value estimates for a state regardless of which trajectory the state is sampled from. Our method, when combined with popular value-based methods, provides improved performance over one-step and multi-step methods on a suite of data-efficient RL benchmarks including MiniGrid, Minatar and Atari100K. We further analyse the reasons for this performance boost through a novel visualisation of the transition graphs of Atari games.
To convey desired behavior to a Reinforcement Learning (RL) agent, a designer must choose a reward function for the environment, arguably the most important knob designers have in interacting with RL agents. Although many reward functions induce the same optimal behavior (Ng et al., 1999), in practice, some of them result in faster learning than others. In this paper, we look at how reward-design choices impact learning speed and seek to identify principles of good reward design that quickly induce target behavior. This reward-identification problem is framed as an optimization problem: Firstly, we advocate choosing state-based rewards that maximize the action gap, making optimal actions easy to distinguish from suboptimal ones. Secondly, we propose minimizing a measure of the horizon, something we call the "subjective discount", over which rewards need to be optimized to encourage agents to make optimal decisions with less lookahead. To solve this optimization problem, we propose a linear-programming based algorithm that efficiently finds a reward function that maximizes action gap and minimizes subjective discount. We test the rewards generated with the algorithm in tabular environments with Q-Learning, and empirically show they lead to faster learning. Although we only focus on Q-Learning because it is perhaps the simplest and most well understood RL algorithm, preliminary results with R-max (Brafman and Tennenholtz, 2000) suggest our results are much more general. Our experiments support three principles of reward design: 1) consistent with existing results, penalizing each step taken induces faster learning than rewarding the goal. 2) When rewarding subgoals along the target trajectory, rewards should gradually increase as the goal gets closer. 3) Dense reward that's nonzero on every state is only good if designed carefully.
Influence maximization (IM) is a combinatorial problem of identifying a subset of nodes called the seed nodes in a network (graph), which when activated, provide a maximal spread of influence in the network for a given diffusion model and a budget for seed set size. IM has numerous applications such as viral marketing, epidemic control, sensor placement and other network-related tasks. However, the uses are limited due to the computational complexity of current algorithms. Recently, learning heuristics for IM have been explored to ease the computational burden. However, there are serious limitations in current approaches such as: (1) IM formulations only consider influence via spread and ignore self activation; (2) scalability to large graphs; (3) generalizability across graph families; (4) low computational efficiency with a large running time to identify seed sets for every test network. In this work, we address each of these limitations through a unique approach that involves (1) formulating a generic IM problem as a Markov decision process that handles both intrinsic and influence activations; (2) employing double Q learning to estimate seed nodes; (3) ensuring scalability via sub-graph based representations; and (4) incorporating generalizability via meta-learning across graph families. Extensive experiments are carried out in various standard networks to validate performance of the proposed Graph Meta Reinforcement learning (GraMeR) framework. The results indicate that GraMeR is multiple orders faster and generic than conventional approaches.
Pedestrian Attribute Recognition (PAR) is a challenging task in intelligent video surveillance. Two key challenges in PAR include complex alignment relations between images and attributes, and imbalanced data distribution. Existing approaches usually formulate PAR as a recognition task. Different from them, this paper addresses it as a decision-making task via a reinforcement learning framework. Specifically, PAR is formulated as a Markov decision process (MDP) by designing ingenious states, action space, reward function and state transition. To alleviate the inter-attribute imbalance problem, we apply an Attribute Grouping Strategy (AGS) by dividing all attributes into subgroups according to their region and category information. Then we employ an agent to recognize each group of attributes, which is trained with Deep Q-learning algorithm. We also propose a Group Optimization Reward (GOR) function to alleviate the intra-attribute imbalance problem. Experimental results on the three benchmark datasets of PETA, RAP and PA100K illustrate the effectiveness and competitiveness of the proposed approach and demonstrate that the application of reinforcement learning to PAR is a valuable research direction.