We study the link between generalization and interference in temporal-difference (TD) learning. Interference is defined as the inner product of two different gradients, representing their alignment. This quantity emerges as being of interest from a variety of observations about neural networks, parameter sharing and the dynamics of learning. We find that TD easily leads to low-interference, under-generalizing parameters, while the effect seems reversed in supervised learning. We hypothesize that the cause can be traced back to the interplay between the dynamics of interference and bootstrapping. This is supported empirically by several observations: the negative relationship between the generalization gap and interference in TD, the negative effect of bootstrapping on interference and the local coherence of targets, and the contrast between the propagation rate of information in TD(0) versus TD($\lambda$) and regression tasks such as Monte-Carlo policy evaluation. We hope that these new findings can guide the future discovery of better bootstrapping methods.

In this work, we investigate the application of Taylor expansions in reinforcement learning. In particular, we propose Taylor expansion policy optimization, a policy optimization formalism that generalizes prior work (e.g., TRPO) as a first-order special case. We also show that Taylor expansions intimately relate to off-policy evaluation. Finally, we show that this new formulation entails modifications which improve the performance of several state-of-the-art distributed algorithms.

Learning-based approaches often outperform hand-coded algorithmic solutions for many problems in robotics. However, learning long-horizon tasks on real robot hardware can be intractable, and transferring a learned policy from simulation to reality is still extremely challenging. We present a novel approach to model-free reinforcement learning that can leverage existing sub-optimal solutions as an algorithmic prior during training and deployment. During training, our gated fusion approach enables the prior to guide the initial stages of exploration, increasing sample-efficiency and enabling learning from sparse long-horizon reward signals. Importantly, the policy can learn to improve beyond the performance of the sub-optimal prior since the prior's influence is annealed gradually. During deployment, the policy's uncertainty provides a reliable strategy for transferring a simulation-trained policy to the real world by falling back to the prior controller in uncertain states. We show the efficacy of our Multiplicative Controller Fusion approach on the task of robot navigation and demonstrate safe transfer from simulation to the real world without any fine tuning. The code for this project is made publicly available at https://sites.google.com/view/mcf-nav/home.

One of the core challenges in the field of control theory and reinforcement learning is adaptive control. It is the problem of controlling dynamical systems when the dynamics of the systems are unknown to the decision-making agents. In adaptive control, agents interact with given systems in order to explore and control them while the long-term objective is to minimize the overall average associated costs. The agent has to balance between exploration and exploitation, learn the dynamics, strategize for further exploration, and exploit the estimation to minimize the overall costs. The sequential nature of agent-system interaction results in challenges in the system identifying, estimation, and control under uncertainty, and these challenges are magnified when the systems are partially observable, i.e. contain hidden underlying dynamics. In the linear systems, when the underlying dynamics are fully observable, the asymptotic optimality of estimation methods has been the topic of study in the last decades [Lai et al., 1982, Lai and Wei, 1987]. Recently, novel techniques and learning algorithms have been developed to study the finite-time behavior of adaptive control algorithms and shed light on the design of optimal methods [Peña et al., 2009, Fiechter, 1997, Abbasi-Yadkori and Szepesvári, 2011]. In particular, Abbasi-Yadkori and Szepesvári [2011] proposes to use the principle of optimism in the face of uncertainty (OFU) to balance exploration and exploitation in LQR, where the state of the system is observable.

Generalization across environments is critical to the successful application of reinforcement learning algorithms to real-world challenges. In this paper, we consider the problem of learning abstractions that generalize in block MDPs, families of environments with a shared latent state space and dynamics structure over that latent space, but varying observations. We leverage tools from causal inference to propose a method of invariant prediction to learn model-irrelevance state abstractions (MISA) that generalize to novel observations in the multi-environment setting. We prove that for certain classes of environments, this approach outputs with high probability a state abstraction corresponding to the causal feature set with respect to the return. We further provide more general bounds on model error and generalization error in the multi-environment setting, in the process showing a connection between causal variable selection and the state abstraction framework for MDPs. We give empirical evidence that our methods work in both linear and nonlinear settings, attaining improved generalization over single- and multi-task baselines.

Options have been shown to be an effective tool in reinforcement learning, facilitating improved exploration and learning. In this paper, we present an approach based on spectral graph theory and derive an algorithm that systematically discovers options without access to a specific reward or task assignment. As opposed to the common practice used in previous methods, our algorithm makes full use of the spectrum of the graph Laplacian. Incorporating modes associated with higher graph frequencies unravels domain subtleties, which are shown to be useful for option discovery. Using geometric and manifold-based analysis, we present a theoretical justification for the algorithm. In addition, we showcase its performance in several domains, demonstrating clear improvements compared to competing methods.

To achieve social interactions within Human-Robot Interaction (HRI) environments is a very challenging task. Most of the current research focuses on Wizard-of-Oz approaches, which neglect the recent development of intelligent robots. On the other hand, real-world scenarios usually do not provide the necessary control and reproducibility which are needed for learning algorithms. In this paper, we propose a virtual simulation environment that implements the Chef's Hat card game, designed to be used in HRI scenarios, to provide a controllable and reproducible scenario for reinforcement-learning algorithms.

This paper is motivated by the following Ad auction problem for an advertiser. An Ad auction is a stochastic game on an Ad exchange platform among a large number of players, the advertisers. In between the time a web user requests a page and the time the page is displayed, usually within a millisecond, a Vickrey-type of second-best-price auction is run to incentivize interested advertisers to bid for an Ad slot to display advertisement. Each advertiser has limited information before each bid: first, her own valuation for a slot depends on an unknown conversion of clicks for the item; secondly, she, should she win the bid, only knows the reward after the user's activities on the website are finished. In addition, she has a budget constraint in this repeated auction. The question is, how should she bid in this online sequential repeated game when there is a large population of bidders competing on the Ad platform, with unknown distributions of the conversion of clicks and rewards? Besides the Ad auction, there are many real-world problems involving a large number of players and unknown systems. Examples include massive multi-player online roleplaying games [30], high frequency tradings [35], and the sharing economy [24].

Sample inefficiency of deep reinforcement learning methods is a major obstacle for their use in real-world applications. In this work, we show how human demonstrations can improve final performance of agents on the Minecraft minigame ObtainDiamond with only 8M frames of environment interaction. We propose a training procedure where policy networks are first trained on human data and later fine-tuned by reinforcement learning. Using a policy exploitation mechanism, experience replay and an additional loss against catastrophic forgetting, our best agent was able to achieve a mean score of 48. Our proposed solution placed 3 rd in the NeurIPS MineRL Competition for Sample-Efficient Reinforcement Learning.

Path-based relational reasoning over knowledge graphs has become increasingly popular due to a variety of downstream applications such as question answering in dialogue systems, fact prediction, and recommender systems. In recent years, reinforcement learning (RL) has provided solutions that are more interpretable and explainable than other deep learning models. However, these solutions still face several challenges, including large action space for the RL agent and accurate representation of entity neighborhood structure. We address these problems by introducing a type-enhanced RL agent that uses the local neighborhood information for efficient path-based reasoning over knowledge graphs. Our solution uses graph neural network (GNN) for encoding the neighborhood information and utilizes entity types to prune the action space. Experiments on real-world dataset show that our method outperforms state-of-the-art RL methods and discovers more novel paths during the training procedure.