Recent advances in multi-agent reinforcement learning (MARL) have achieved super-human performance in games like Quake 3 and Dota 2. Unfortunately, these techniques require orders-of-magnitude more training rounds than humans and don't generalize to new agent configurations even on the same game. In this work, we propose Collaborative Q-learning (CollaQ) that achieves state-of-the-art performance in the StarCraft multi-agent challenge and supports ad hoc team play. We first formulate multi-agent collaboration as a joint optimization on reward assignment and show that each agent has an approximately optimal policy that decomposes into two parts: one part that only relies on the agent's own state, and the other part that is related to states of nearby agents. Following this novel finding, CollaQ decomposes the Q-function of each agent into a self term and an interactive term, with a Multi-Agent Reward Attribution (MARA) loss that regularizes the training. CollaQ is evaluated on various StarCraft maps and shows that it outperforms existing state-of-the-art techniques (i.e., QMIX, QTRAN, and VDN) by improving the win rate by 40% with the same number of samples. In the more challenging ad hoc team play setting (i.e., reweight/add/remove units without re-training or finetuning), CollaQ outperforms previous SoTA by over 30%.
Epidemiologists model the dynamics of epidemics in order to propose control strategies based on pharmaceutical and non-pharmaceutical interventions (contact limitation, lock down, vaccination, etc). Hand-designing such strategies is not trivial because of the number of possible interventions and the difficulty to predict long-term effects. This task can be cast as an optimization problem where state-of-the-art machine learning algorithms such as deep reinforcement learning, might bring significant value. However, the specificity of each domain -- epidemic modelling or solving optimization problem -- requires strong collaborations between researchers from different fields of expertise. This is why we introduce EpidemiOptim, a Python toolbox that facilitates collaborations between researchers in epidemiology and optimization. EpidemiOptim turns epidemiological models and cost functions into optimization problems via a standard interface commonly used by optimization practitioners (OpenAI Gym). Reinforcement learning algorithms based on Q-Learning with deep neural networks (DQN) and evolutionary algorithms (NSGA-II) are already implemented. We illustrate the use of EpidemiOptim to find optimal policies for dynamical on-off lock-down control under the optimization of death toll and economic recess using a Susceptible-Exposed-Infectious-Removed (SEIR) model for COVID-19. Using EpidemiOptim and its interactive visualization platform in Jupyter notebooks, epidemiologists, optimization practitioners and others (e.g. economists) can easily compare epidemiological models, costs functions and optimization algorithms to address important choices to be made by health decision-makers.
Unsupervised methods are promising for abstractive text summarization in that the parallel corpora is not required. However, their performance is still far from being satisfied, therefore research on promising solutions is on-going. In this paper, we propose a new approach based on Q-learning with an edit-based summarization. The method combines two key modules to form an Editorial Agent and Language Model converter (EALM). The agent predicts edit actions (e.t., delete, keep, and replace), and then the LM converter deterministically generates a summary on the basis of the action signals. Q-learning is leveraged to train the agent to produce proper edit actions. Experimental results show that EALM delivered competitive performance compared with the previous encoder-decoder-based methods, even with truly zero paired data (i.e., no validation set). Defining the task as Q-learning enables us not only to develop a competitive method but also to make the latest techniques in reinforcement learning available for unsupervised summarization. We also conduct qualitative analysis, providing insights into future study on unsupervised summarizers.
Evolution strategies (ES), as a family of black-box optimization algorithms, recently emerge as a scalable alternative to reinforcement learning (RL) approaches such as Q-learning or policy gradient, and are much faster when many central processing units (CPUs) are available due to better parallelization. In this paper, we propose a systematic incremental learning method for ES in dynamic environments. The goal is to adjust previously learned policy to a new one incrementally whenever the environment changes. We incorporate an instance weighting mechanism with ES to facilitate its learning adaptation, while retaining scalability of ES. During parameter updating, higher weights are assigned to instances that contain more new knowledge, thus encouraging the search distribution to move towards new promising areas of parameter space. We propose two easy-to-implement metrics to calculate the weights: instance novelty and instance quality. Instance novelty measures an instance's difference from the previous optimum in the original environment, while instance quality corresponds to how well an instance performs in the new environment. The resulting algorithm, Instance Weighted Incremental Evolution Strategies (IW-IES), is verified to achieve significantly improved performance on challenging RL tasks ranging from robot navigation to locomotion. This paper thus introduces a family of scalable ES algorithms for RL domains that enables rapid learning adaptation to dynamic environments.
We show that Reinforcement Learning (RL) methods for solving Text-Based Games (TBGs) often fail to generalize on unseen games, especially in small data regimes. To address this issue, we propose Context Relevant Episodic State Truncation (CREST) for irrelevant token removal in observation text for improved generalization. Our method first trains a base model using Q-learning, which typically overfits the training games. The base model's action token distribution is used to perform observation pruning that removes irrelevant tokens. A second bootstrapped model is then retrained on the pruned observation text. Our bootstrapped agent shows improved generalization in solving unseen TextWorld games, using 10x-20x fewer training games compared to previous state-of-the-art methods despite requiring less number of training episodes.
Deep active inference has been proposed as a scalable approach to perception and action that deals with large policy and state spaces. However, current models are limited to fully observable domains. In this paper, we describe a deep active inference model that can learn successful policies directly from high-dimensional sensory inputs. The deep learning architecture optimizes a variant of the expected free energy and encodes the continuous state representation by means of a variational autoencoder. We show, in the OpenAI benchmark, that our approach has comparable or better performance than deep Q-learning, a state-of-the-art deep reinforcement learning algorithm.
It is common for people to create different types of charts to explore a multi-dimensional dataset (table). However, to recommend commonly composed charts in real world, one should take the challenges of efficiency, imbalanced data and table context into consideration. In this paper, we propose Table2Charts framework which learns common patterns from a large corpus of (table, charts) pairs. Based on deep Q-learning with copying mechanism and heuristic searching, Table2Charts does table-to-sequence generation, where each sequence follows a chart template. On a large spreadsheet corpus with 165k tables and 266k charts, we show that Table2Charts could learn a shared representation of table fields so that recommendation tasks on different chart types could mutually enhance each other. Table2Charts outperforms other chart recommendation systems in both multi-type task (with doubled recall numbers R@3=0.61 and R@1=0.43) and human evaluations.
Recent studies have shown that deep reinforcement learning agents are vulnerable to small adversarial perturbations on the agent's inputs, which raises concerns about deploying such agents in the real world. To address this issue, we propose RADIAL-RL, a principled framework to train reinforcement learning agents with improved robustness against $l_p$-norm bounded adversarial attacks. Our framework is compatible with popular deep reinforcement learning algorithms and we demonstrate its performance with deep Q-learning, A3C and PPO. We experiment on three deep RL benchmarks (Atari, MuJoCo and ProcGen) to show the effectiveness of our robust training algorithm. Our RADIAL-RL agents consistently outperform prior methods when tested against attacks of varying strength and are more computationally efficient to train. In addition, we propose a new evaluation method called Greedy Worst-Case Reward (GWC) to measure attack agnostic robustness of deep RL agents. We show that GWC can be evaluated efficiently and is a good estimate of the reward under the worst possible sequence of adversarial attacks. All code used for our experiments is available at https://github.com/tuomaso/radial_rl_v2.
We explore value-based multi-agent reinforcement learning (MARL) in the popular paradigm of centralized training with decentralized execution (CTDE). CTDE has an important concept, Individual-Global-Max (IGM) principle, which requires the consistency between joint and local action selections to support efficient local decision-making. However, in order to achieve scalability, existing MARL methods either limit representation expressiveness of their value function classes or relax the IGM consistency, which may suffer from instability risk or may not perform well in complex domains. This paper presents a novel MARL approach, called duPLEX dueling multi-agent Q-learning (QPLEX), which takes a duplex dueling network architecture to factorize the joint value function. This duplex dueling structure encodes the IGM principle into the neural network architecture and thus enables efficient value function learning. Theoretical analysis shows that QPLEX achieves a complete IGM function class. Empirical experiments on StarCraft II micromanagement tasks demonstrate that QPLEX significantly outperforms state-of-the-art baselines in both online and offline data collection settings, and also reveal that QPLEX achieves high sample efficiency and can benefit from offline datasets without additional online exploration.
This paper introduces single-partition adaptive Q-learning (SPAQL), an algorithm for model-free episodic reinforcement learning (RL), which adaptively partitions the state-action space of a Markov decision process (MDP), while simultaneously learning a time-invariant policy (i. e., the mapping from states to actions does not depend explicitly on the episode time step) for maximizing the cumulative reward. The trade-off between exploration and exploitation is handled by using a mixture of upper confidence bounds (UCB) and Boltzmann exploration during training, with a temperature parameter that is automatically tuned as training progresses. The algorithm is an improvement over adaptive Q-learning (AQL). It converges faster to the optimal solution, while also using fewer arms. Tests on episodes with a large number of time steps show that SPAQL has no problems scaling, unlike AQL. Based on this empirical evidence, we claim that SPAQL may have a higher sample efficiency than AQL, thus being a relevant contribution to the field of efficient model-free RL methods.
Experience replay is central to off-policy algorithms in deep reinforcement learning (RL), but there remain significant gaps in our understanding. We therefore present a systematic and extensive analysis of experience replay in Q-learning methods, focusing on two fundamental properties: the replay capacity and the ratio of learning updates to experience collected (replay ratio). Our additive and ablative studies upend conventional wisdom around experience replay -- greater capacity is found to substantially increase the performance of certain algorithms, while leaving others unaffected. Counterintuitively we show that theoretically ungrounded, uncorrected n-step returns are uniquely beneficial while other techniques confer limited benefit for sifting through larger memory. Separately, by directly controlling the replay ratio we contextualize previous observations in the literature and empirically measure its importance across a variety of deep RL algorithms. Finally, we conclude by testing a set of hypotheses on the nature of these performance benefits.
In this paper, we establish a theoretical comparison between the asymptotic mean-squared error of Double Q-learning and Q-learning. Our result builds upon an analysis for linear stochastic approximation based on Lyapunov equations and applies to both tabular setting and with linear function approximation, provided that the optimal policy is unique and the algorithms converge. We show that the asymptotic mean-squared error of Double Q-learning is exactly equal to that of Q-learning if Double Q-learning uses twice the learning rate of Q-learning and outputs the average of its two estimators. We also present some practical implications of this theoretical observation using simulations.
In Reinforcement Learning (RL), Convolutional Neural Networks(CNNs) have been successfully applied as function approximators in Deep Q-Learning algorithms, which seek to learn action-value functions and policies in various environments. However, to date, there has been little work on the learning of symmetry-transformation equivariant representations of the input environment state. In this paper, we propose the use of Equivariant CNNs to train RL agents and study their inductive bias for transformation equivariant Q-value approximation. We demonstrate that equivariant architectures can dramatically enhance the performance and sample efficiency of RL agents in a highly symmetric environment while requiring fewer parameters. Additionally, we show that they are robust to changes in the environment caused by affine transformations.
Off-policy deep reinforcement learning (RL) has been successful in a range of challenging domains. However, standard off-policy RL algorithms can suffer from several issues, such as instability in Q-learning and balancing exploration and exploitation. To mitigate these issues, we present SUNRISE, a simple unified ensemble method, which is compatible with various off-policy RL algorithms. SUNRISE integrates two key ingredients: (a) ensemble-based weighted Bellman backups, which re-weight target Q-values based on uncertainty estimates from a Q-ensemble, and (b) an inference method that selects actions using the highest upper-confidence bounds for efficient exploration. By enforcing the diversity between agents using Bootstrap with random initialization, we show that these different ideas are largely orthogonal and can be fruitfully integrated, together further improving the performance of existing off-policy RL algorithms, such as Soft Actor-Critic and Rainbow DQN, for both continuous and discrete control tasks on both low-dimensional and high-dimensional environments. Our training code is available at https://github.com/pokaxpoka/sunrise.
In this paper, we study collaborative filtering in an interactive setting, in which the recommender agents iterate between making recommendations and updating the user profile based on the interactive feedback. The most challenging problem in this scenario is how to suggest items when the user profile has not been well established, i.e., recommend for cold-start users or warm-start users with taste drifting. Existing approaches either rely on overly pessimistic linear exploration strategy or adopt meta-learning based algorithms in a full exploitation way. In this work, to quickly catch up with the user's interests, we propose to represent the exploration policy with a neural network and directly learn it from the feedback data. Specifically, the exploration policy is encoded in the weights of multi-channel stacked self-attention neural networks and trained with efficient Q-learning by maximizing users' overall satisfaction in the recommender systems. The key insight is that the satisfied recommendations triggered by the exploration recommendation can be viewed as the exploration bonus (delayed reward) for its contribution on improving the quality of the user profile. Therefore, the proposed exploration policy, to balance between learning the user profile and making accurate recommendations, can be directly optimized by maximizing users' long-term satisfaction with reinforcement learning. Extensive experiments and analysis conducted on three benchmark collaborative filtering datasets have demonstrated the advantage of our method over state-of-the-art methods.
It is still common to use Q-learning and temporal difference (TD) learning-even though they have divergence issues and sound Gradient TD alternatives exist-because divergence seems rare and they typically perform well. However, recent work with large neural network learning systems reveals that instability is more common than previously thought. Practitioners face a difficult dilemma: choose an easy to use and performant TD method, or a more complex algorithm that is more sound but harder to tune and all but unexplored with non-linear function approximation or control. In this paper, we introduce a new method called TD with Regularized Corrections (TDRC), that attempts to balance ease of use, soundness, and performance. It behaves as well as TD, when TD performs well, but is sound in cases where TD diverges. We empirically investigate TDRC across a range of problems, for both prediction and control, and for both linear and non-linear function approximation, and show, potentially for the first time, that gradient TD methods could be a better alternative to TD and Q-learning.
Recent results in supervised learning suggest that while overparameterized models have the capacity to overfit, they in fact generalize quite well. We ask whether the same phenomenon occurs for offline contextual bandits. Our results are mixed. Value-based algorithms benefit from the same generalization behavior as overparameterized supervised learning, but policy-based algorithms do not. We show that this discrepancy is due to the \emph{action-stability} of their objectives. An objective is action-stable if there exists a prediction (action-value vector or action distribution) which is optimal no matter which action is observed. While value-based objectives are action-stable, policy-based objectives are unstable. We formally prove upper bounds on the regret of overparameterized value-based learning and lower bounds on the regret for policy-based algorithms. In our experiments with large neural networks, this gap between action-stable value-based objectives and unstable policy-based objectives leads to significant performance differences.
We introduce the lookahead-bounded Q-learning (LBQL) algorithm, a new, provably convergent variant of Q-learning that seeks to improve the performance of standard Q-learning in stochastic environments through the use of ``lookahead'' upper and lower bounds. To do this, LBQL employs previously collected experience and each iteration's state-action values as dual feasible penalties to construct a sequence of sampled information relaxation problems. The solutions to these problems provide estimated upper and lower bounds on the optimal value, which we track via stochastic approximation. These quantities are then used to constrain the iterates to stay within the bounds at every iteration. Numerical experiments on benchmark problems show that LBQL exhibits faster convergence and more robustness to hyperparameters when compared to standard Q-learning and several related techniques. Our approach is particularly appealing in problems that require expensive simulations or real-world interactions.
QMIX is a popular $Q$-learning algorithm for cooperative MARL in the centralised training and decentralised execution paradigm. In order to enable easy decentralisation, QMIX restricts the joint action $Q$-values it can represent to be a monotonic mixing of each agent's utilities. However, this restriction prevents it from representing value functions in which an agent's ordering over its actions can depend on other agents' actions. To analyse this representational limitation, we first formalise the objective QMIX optimises, which allows us to view QMIX as an operator that first computes the $Q$-learning targets and then projects them into the space representable by QMIX. This projection returns a representable $Q$-value that minimises the unweighted squared error across all joint actions. We show in particular that this projection can fail to recover the optimal policy even with access to $Q^*$, which primarily stems from the equal weighting placed on each joint action. We rectify this by introducing a weighting into the projection, in order to place more importance on the better joint actions. We propose two weighting schemes and prove that they recover the correct maximal action for any joint action $Q$-values, and therefore for $Q^*$ as well. Based on our analysis and results in the tabular setting, we introduce two scalable versions of our algorithm, Centrally-Weighted (CW) QMIX and Optimistically-Weighted (OW) QMIX and demonstrate improved performance on both predator-prey and challenging multi-agent StarCraft benchmark tasks.
Semantic cues and statistical regularities in real-world environment layouts can improve efficiency for navigation in novel environments. This paper learns and leverages such semantic cues for navigating to objects of interest in novel environments, by simply watching YouTube videos. This is challenging because YouTube videos don't come with labels for actions or goals, and may not even showcase optimal behavior. Our method tackles these challenges through the use of Q-learning on pseudo-labeled transition quadruples (image, action, next image, reward). We show that such off-policy Q-learning from passive data is able to learn meaningful semantic cues for navigation. These cues, when used in a hierarchical navigation policy, lead to improved efficiency at the ObjectGoal task in visually realistic simulations. We observe a relative improvement of 15-83% over end-to-end RL, behavior cloning, and classical methods, while using minimal direct interaction.