Off-policy learning allows us to learn about possible policies of behavior from experience generated by a different behavior policy. Temporal difference (TD) learning algorithms can become unstable when combined with function approximation and off-policy sampling - this is known as the ''deadly triad''. Emphatic temporal difference (ETD($\lambda$)) algorithm ensures convergence in the linear case by appropriately weighting the TD($\lambda$) updates. In this paper, we extend the use of emphatic methods to deep reinforcement learning agents. We show that naively adapting ETD($\lambda$) to popular deep reinforcement learning algorithms, which use forward view multi-step returns, results in poor performance. We then derive new emphatic algorithms for use in the context of such algorithms, and we demonstrate that they provide noticeable benefits in small problems designed to highlight the instability of TD methods. Finally, we observed improved performance when applying these algorithms at scale on classic Atari games from the Arcade Learning Environment.

Reinforcement learning (RL) has gained increasing interest since the demonstration it was able to reach human performance on video game benchmarks using deep Q-learning (DQN). The current consensus for training neural networks on such complex environments is to rely on gradient-based optimization. Although alternative Bayesian deep learning methods exist, most of them still rely on gradient-based optimization, and they typically do not scale on benchmarks such as the Atari game environment. Moreover none of these approaches allow performing the analytical inference for the weights and biases defining the neural network. In this paper, we present how we can adapt the temporal difference Q-learning framework to make it compatible with the tractable approximate Gaussian inference (TAGI), which allows learning the parameters of a neural network using a closed-form analytical method. Throughout the experiments with on- and off-policy reinforcement learning approaches, we demonstrate that TAGI can reach a performance comparable to backpropagation-trained networks while using fewer hyperparameters, and without relying on gradient-based optimization.

Reinforcement Learning (RL) has achieved several remarkable empirical successes in the last decade, which include playing Atari 2600 video games at superhuman levels (Mnih et al., 2015), AlphaGo or AlphaGo Zero surpassing champions in Go (Silver et al., 2018), AlphaStar's victory over top-ranked professional players in StarCraft (Vinyals et al., 2019), or practical self-driving cars. These applications all correspond to the setting of rich observations, where the state space is very large and where observations may be images, text or audio data. In contrast, most provably efficient RL algorithms are still limited to the classical tabular setting where the state space is small (Kearns and Singh, 2002; Brafman and Tennenholtz, 2002; Azar et al., 2017; Dann et al., 2019) and do not scale to the rich observation setting. To derive guarantees for large state spaces, much of the existing work in RL theory relies on a realizability and a low-rank assumption (Krishnamurthy et al., 2016; Jiang et al., 2017; Dann et al., 2018; Du et al., 2019a; Misra et al., 2020; Agarwal et al., 2020b). Different notions of rank have been adopted in the literature, including that of a low-rank transition matrix (Jin et al., 2020a), a low Bellman rank (Jiang et al., 2017), Wittness rank (Sun et al., 2019), Eluder dimension (Osband and Van Roy, 2014), Bellman-Eluder dimension (Jin et al., 2021), or bilinear classes (Du et al., 2021).

Multi-Agent Path Finding (MAPF) is essential to large-scale robotic systems. Recent methods have applied reinforcement learning (RL) to learn decentralized polices in partially observable environments. A fundamental challenge of obtaining collision-free policy is that agents need to learn cooperation to handle congested situations. This paper combines communication with deep Q-learning to provide a novel learning based method for MAPF, where agents achieve cooperation via graph convolution. To guide RL algorithm on long-horizon goal-oriented tasks, we embed the potential choices of shortest paths from single source as heuristic guidance instead of using a specific path as in most existing works. Our method treats each agent independently and trains the model from a single agent's perspective. The final trained policy is applied to each agent for decentralized execution. The whole system is distributed during training and is trained under a curriculum learning strategy. Empirical evaluation in obstacle-rich environment indicates the high success rate with low average step of our method.

Involving humans directly for the benefit of AI agents' training is getting traction thanks to several advances in reinforcement learning and human-in-the-loop learning. Humans can provide rewards to the agent, demonstrate tasks, design a curriculum, or act in the environment, but these benefits also come with architectural, functional design and engineering complexities. We present Cogment, a unifying open-source framework that introduces an actor formalism to support a variety of humans-agents collaboration typologies and training approaches. It is also scalable out of the box thanks to a distributed micro service architecture, and offers solutions to the aforementioned complexities.

Unsupervised domain adaptation is used in many machine learning applications where, during training, a model has access to unlabeled data in the target domain, and a related labeled dataset. In this paper, we introduce a novel and general domain-adversarial framework. Specifically, we derive a novel generalization bound for domain adaptation that exploits a new measure of discrepancy between distributions based on a variational characterization of f-divergences. It recovers the theoretical results from Ben-David et al. (2010a) as a special case and supports divergences used in practice. Based on this bound, we derive a new algorithmic framework that introduces a key correction in the original adversarial training method of Ganin et al. (2016). We show that many regularizers and ad-hoc objectives introduced over the last years in this framework are then not required to achieve performance comparable to (if not better than) state-of-the-art domain-adversarial methods. Experimental analysis conducted on real-world natural language and computer vision datasets show that our framework outperforms existing baselines, and obtains the best results for f-divergences that were not considered previously in domain-adversarial learning.

Recent years have witnessed great success in handling node classification tasks with Graph Neural Networks (GNNs). However, most existing GNNs are based on the assumption that node samples for different classes are balanced, while for many real-world graphs, there exists the problem of class imbalance, i.e., some classes may have much fewer samples than others. In this case, directly training a GNN classifier with raw data would under-represent samples from those minority classes and result in sub-optimal performance. This paper presents GraphMixup, a novel mixup-based framework for improving class-imbalanced node classification on graphs. However, directly performing mixup in the input space or embedding space may produce out-of-domain samples due to the extreme sparsity of minority classes; hence we construct semantic relation spaces that allows the Feature Mixup to be performed at the semantic level. Moreover, we apply two context-based self-supervised techniques to capture both local and global information in the graph structure and then propose Edge Mixup specifically for graph data. Finally, we develop a \emph{Reinforcement Mixup} mechanism to adaptively determine how many samples are to be generated by mixup for those minority classes. Extensive experiments on three real-world datasets show that GraphMixup yields truly encouraging results for class-imbalanced node classification tasks.

Offline reinforcement learning (RL) tries to learn the near-optimal policy with recorded offline experience without online exploration. Current offline RL research includes: 1) generative modeling, i.e., approximating a policy using fixed data; and 2) learning the state-action value function. While most research focuses on the state-action function part through reducing the bootstrapping error in value function approximation induced by the distribution shift of training data, the effects of error propagation in generative modeling have been neglected. In this paper, we analyze the error in generative modeling. We propose AQL (action-conditioned Q-learning), a residual generative model to reduce policy approximation error for offline RL. We show that our method can learn more accurate policy approximations in different benchmark datasets. In addition, we show that the proposed offline RL method can learn more competitive AI agents in complex control tasks under the multiplayer online battle arena (MOBA) game Honor of Kings.

Ever since DeepMind published its paper "Playing Atari with Deep Reinforcement Learning", many promising results have come out, with perhaps the most famous one being AlphaGo. Before we can understand how these models work, however, we need to understand some basic principles of reinforcement learning. I think the best introduction to these concepts is through a much simpler problem -- the so-called "k-armed bandit problem". First, let's define what the k-armed bandit problem is. The phrase "k-armed bandit" conjures up images of Boba Fett from Star Wars crossed with an octopus, but fortunately things are not that crazy.

We consider the offline reinforcement learning (RL) setting where the agent aims to optimize the policy solely from the data without further environment interactions. In offline RL, the distributional shift becomes the primary source of difficulty, which arises from the deviation of the target policy being optimized from the behavior policy used for data collection. This typically causes overestimation of action values, which poses severe problems for model-free algorithms that use bootstrapping. To mitigate the problem, prior offline RL algorithms often used sophisticated techniques that encourage underestimation of action values, which introduces an additional set of hyperparameters that need to be tuned properly. In this paper, we present an offline RL algorithm that prevents overestimation in a more principled way. Our algorithm, OptiDICE, directly estimates the stationary distribution corrections of the optimal policy and does not rely on policy-gradients, unlike previous offline RL algorithms. Using an extensive set of benchmark datasets for offline RL, we show that OptiDICE performs competitively with the state-of-the-art methods.