Papers

Reinforcement Learning (RL) agents in the real world must satisfy safety constraints in addition to maximizing a reward objective. Model-based RL algorithms hold promise for reducing unsafe real-world actions: they may synthesize policies that obey all constraints using simulated samples from a learned model. However, imperfect models can result in real-world constraint violations even for actions that are predicted to satisfy all constraints. We propose Conservative and Adaptive Penalty (CAP), a model-based safe RL framework that accounts for potential modeling errors by capturing model uncertainty and adaptively exploiting it to balance the reward and the cost objectives. First, CAP inflates predicted costs using an uncertainty-based penalty. Theoretically, we show that policies that satisfy this conservative cost constraint are guaranteed to also be feasible in the true environment. We further show that this guarantees the safety of all intermediate solutions during RL training. Further, CAP adaptively tunes this penalty during training using true cost feedback from the environment. We evaluate this conservative and adaptive penalty-based approach for model-based safe RL extensively on state and image-based environments. Our results demonstrate substantial gains in sample-efficiency while incurring fewer violations than prior safe RL algorithms. Code is available at: https://github.com/Redrew/CAP

poster session 1 @ blue 1, poster session 8 @ blue 1, poster session 1, poster session 8

When designing algorithms for finite-time-horizon episodic reinforcement learning problems, a common approach is to introduce a fictitious discount factor and use stationary policies for approximations. Empirically, it has been shown that the fictitious discount factor helps reduce variance, and stationary policies serve to save the per-iteration computational cost. Theoretically, however, there is no existing work on convergence analysis for algorithms with this fictitious discount recipe. This paper takes the first step towards analyzing these algorithms. It focuses on two vanilla policy gradient (VPG) variants: the first being a widely used variant with discounted advantage estimations (DAE), the second with an additional fictitious discount factor in the score functions of the policy gradient estimators. Non-asymptotic convergence guarantees are established for both algorithms, and the additional discount factor is shown to reduce the bias introduced in DAE and thus improve the algorithm convergence asymptotically. A key ingredient of our analysis is to connect three settings of Markov decision processes (MDPs): the finite-time-horizon, the average reward and the discounted settings. To our best knowledge, this is the first theoretical guarantee on fictitious discount algorithms for the episodic reinforcement learning of finite-time-horizon MDPs, which also leads to the (first) global convergence of policy gradient methods for finite-time-horizon episodic reinforcement learning.

poster session 1 @ blue 1, poster session 8 @ blue 1, poster session 1, poster session 8

We study regret minimization in finite horizon tabular Markov decision processes (MDPs) under the constraints of differential privacy (DP). This is motivated by the widespread applications of reinforcement learning (RL) in real-world sequential decision making problems, where protecting users' sensitive and private information is becoming paramount. We consider two variants of DP -- joint DP (JDP), where a centralized agent is responsible for protecting users' sensitive data and local DP (LDP), where information needs to be protected directly on the user side. We first propose two general frameworks -- one for policy optimization and another for value iteration -- for designing private, optimistic RL algorithms. We then instantiate these frameworks with suitable privacy mechanisms to satisfy JDP and LDP requirements, and simultaneously obtain sublinear regret guarantees. The regret bounds show that under JDP, the cost of privacy is only a lower order additive term, while for a stronger privacy protection under LDP, the cost suffered is multiplicative. Finally, the regret bounds are obtained by a unified analysis, which, we believe, can be extended beyond tabular MDPs.

poster session 1 @ blue 1, poster session 8 @ blue 1, oral session 1 @ blue 1, poster session 1, poster session 8, oral session 1

We study reward design strategies for incentivizing a reinforcement learning agent to adopt a policy from a set of admissible policies. The goal of the reward designer is to modify the underlying reward function cost-efficiently while ensuring that any approximately optimal deterministic policy under the new reward function is admissible and performs well under the original reward function. This problem can be viewed as a dual to the problem of optimal reward poisoning attacks: instead of forcing an agent to adopt a specific policy, the reward designer incentivizes an agent to avoid taking actions that are inadmissible in certain states. Perhaps surprisingly, and in contrast to the problem of optimal reward poisoning attacks, we first show that the reward design problem for admissible policy teaching is computationally challenging, and it is NP-hard to find an approximately optimal reward modification. We then proceed by formulating a surrogate problem whose optimal solution approximates the optimal solution to the reward design problem in our setting, but is more amenable to optimization techniques and analysis. For this surrogate problem, we present characterization results that provide bounds on the value of the optimal solution. Finally, we design a local search algorithm to solve the surrogate problem and showcase its utility using simulation-based experiments.

poster session 1 @ blue 1, poster session 8 @ blue 1, oral session 1 @ blue 1, poster session 1, poster session 8, oral session 1

Online reinforcement learning (RL) algorithms are often difficult to deploy in complex human-facing applications as they may learn slowly and have poor early performance. To address this, we introduce a practical algorithm for incorporating human insight to speed learning. Our algorithm, Constraint Sampling Reinforcement Learning (CSRL), incorporates prior domain knowledge as constraints/restrictions on the RL policy. It takes in multiple potential policy constraints to maintain robustness to misspecification of individual constraints while leveraging helpful ones to learn quickly. Given a base RL learning algorithm (ex. UCRL, DQN, Rainbow) we propose an upper confidence with elimination scheme that leverages the relationship between the constraints, and their observed performance, to adaptively switch among them. We instantiate our algorithm with DQN-type algorithms and UCRL as base algorithms, and evaluate our algorithm in four environments, including three simulators based on real data: recommendations, educational activity sequencing, and HIV treatment sequencing. In all cases, CSRL learns a good policy faster than baselines.

poster session 1 @ blue 1, poster session 8 @ blue 1, poster session 1, poster session 8

Simplicial neural networks (SNNs) have recently emerged as a new direction in graph learning which expands the idea of convolutional architectures from node space to simplicial complexes on graphs. Instead of predominantly assessing pairwise relations among nodes as in the current practice, simplicial complexes allow us to describe higher-order interactions and multi-node graph structures. By building upon connection between the convolution operation and the new block Hodge-Laplacian, we propose the first SNN for link prediction. Our new Block Simplicial Complex Neural Networks (BScNets) model generalizes existing graph convolutional network (GCN) frameworks by systematically incorporating salient interactions among multiple higher-order graph structures of different dimensions. We discuss theoretical foundations behind BScNets and illustrate its utility for link prediction on eight real-world and synthetic datasets. Our experiments indicate that BScNets outperforms the state-of-the-art models by a significant margin while maintaining low computation costs. Finally, we show utility of BScNets as a new promising alternative for tracking spread of infectious diseases such as COVID-19 and measuring the effectiveness of the healthcare risk mitigation strategies.

poster session 1 @ blue 1, poster session 8 @ blue 1, poster session 1, poster session 8

Meta reinforcement learning (meta-RL) aims to learn a policy solving a set of training tasks simultaneously and quickly adapting to new tasks. It requires massive amounts of data drawn from training tasks to infer the common structure shared among tasks. Without heavy reward engineering, the sparse rewards in long-horizon tasks exacerbate the problem of sample efficiency in meta-RL. Another challenge in meta-RL is the discrepancy of difficulty level among tasks, which might cause one easy task dominating learning of the shared policy and thus preclude policy adaptation to new tasks. This work introduces a novel objective function to learn an action translator among training tasks. We theoretically verify that the value of the transferred policy with the action translator can be close to the value of the source policy and our objective function (approximately) upper bounds the value difference. We propose to combine the action translator with context-based meta-RL algorithms for better data collection and moreefficient exploration during meta-training. Our approach em-pirically improves the sample efficiency and performance ofmeta-RL algorithms on sparse-reward tasks.

poster session 1 @ blue 1, poster session 8 @ blue 1, poster session 1, poster session 8

This work explores how to learn robust and generalizable state representation from image-based observations with deep reinforcement learning methods. Addressing the computational complexity, stringent assumptions, and representation collapse challenges in existing work of bisimulation metric, we devise Simple State Representation (SimSR) operator, which achieves equivalent functionality while reducing the complexity by an order in comparison with bisimulation metric. SimSR enables us to design a stochastic-approximation-based method that can practically learn the mapping functions (encoders) from observations to latent representation space. Besides the theoretical analysis, we experimented and compared our work with recent state-of-the-art solutions in visual MuJoCo tasks. The results shows that our model generally achieves better performance and has better robustness and good generalization.

poster session 1 @ blue 1, poster session 8 @ blue 1, oral session 8 @ blue 1, poster session 1, poster session 8, oral session 8

Policy gradient methods have been frequently applied to problems in control and reinforcement learning with great success, yet existing convergence analysis still relies on non-intuitive, impractical and often opaque conditions. In particular, existing rates are achieved in limited settings, under strict smoothness and bounded conditions. In this work, we establish explicit convergence rates of policy gradient methods without relying on these conditions, instead extending the convergence regime to weakly smooth policy classes with L2 integrable gradient. We provide intuitive examples to illustrate the insight behind these new conditions. We also characterize the sufficiency conditions for the ergodicity of near-linear MDPs, which represent an important class of problems. Notably, our analysis also shows that fast convergence rates are achievable for both the standard policy gradient and the natural policy gradient algorithms under these assumptions. Lastly we provide conditions and analysis for optimality of the converged policies.

poster session 1 @ blue 1, poster session 8 @ blue 1, poster session 1, poster session 8

Reinforcement learning is widely used in applications where one needs to perform sequential decisions while interacting with the environment. The problem becomes more challenging when the decision requirement includes satisfying some safety constraints. The problem is mathematically formulated as constrained Markov decision process (CMDP). In the literature, various algorithms are available to solve CMDP problems in a model-free manner to achieve $\epsilon$-optimal cumulative reward with $\epsilon$ feasible policies. An $\epsilon$-feasible policy implies that it suffers from constraint violation. An important question here is whether we can achieve $\epsilon$-optimal cumulative reward with zero constraint violations or not. To achieve that, we advocate the use of randomized primal-dual approach to solve the CMDP problems and propose a conservative stochastic primal-dual algorithm (CSPDA) which is shown to exhibit $\tilde{\mathcal{O}}\left(1/\epsilon^2\right)$ sample complexity to achieve $\epsilon$-optimal cumulative reward with zero constraint violations. In the prior works, the best available sample complexity for the $\epsilon$-optimal policy with zero constraint violation is $\tilde{\mathcal{O}}\left(1/\epsilon^5\right)$. Hence, the proposed algorithm provides a significant improvement as compared to the state of the art.

poster session 1 @ blue 1, poster session 8 @ blue 1, poster session 1, poster session 8

Real world tasks are hierarchical and compositional. Tasks can be composed of multiple subtasks (or sub-goals) that are dependent on each other. These subtasks are defined in terms of entities (e.g., "apple", "pear") that can be recombined to form new subtasks (e.g., "pickup apple", and "pickup pear"). To solve these tasks efficiently, an agent must infer subtask dependencies (e.g. an agent must execute "pickup apple" before "place apple in pot"), and generalize the inferred dependencies to new subtasks (e.g. "place apple in pot" is similar to "place apple in pan"). Moreover, an agent may also need to solve unseen tasks, which can involve unseen entities. To this end, we formulate parameterized subtask graph inference (PSGI), a method for modeling subtask dependencies using first-order logic with factored entities. To facilitate this, we learn parameter attributes in a zero-shot manner, which are used as quantifiers (e.g. is_pickable(X)) for the factored subtask graph. We show this approach accurately learns the latent structure on hierarchical and compositional tasks more efficiently than prior work, and show PSGI can generalize by modelling structure on subtasks unseen during adaptation.

poster session 1 @ blue 1, poster session 8 @ blue 1, oral session 1 @ blue 1, poster session 1, poster session 8, oral session 1

Contrastive learning has been widely applied to graph representation learning, where the view generators play a vital role in generating effective contrastive samples. Most of the existing contrastive learning methods employ pre-defined view generation methods, e.g., node drop or edge perturbation, which are independent of the input data and cannot well maintain the original semantic structures. To address this issue, we propose a novel framework named Automated Graph Contrastive Learning (AutoGCL) in this paper. The core of AutoGCL is a set of learnable graph view generators that learn a probability distribution over the contrastive samples conditioned on the input graphs. It is able to preserve the most discriminative structures of the original graphs as well as providing adequate augmentation variances for contrastive learning. Moreover, we propose a joint training strategy to train the learnable view generators, the graph encoder, and the classifier in an end-to-end manner, forcing the generations of the view generators to be topologically different but semantically similar. Extensive experiments on semi-supervised learning, unsupervised learning, and transfer learning demonstrate that our AutoGCL framework is superior to the state-of-the-art graph contrastive learning methods. In addition, the visualization results show that the learnable view generators can deliver more compact and semantically meaningful contrastive samples compared against the existing view generation methods.

poster session 1 @ blue 1, poster session 8 @ blue 1, poster session 1, poster session 8

We posit a new mechanism for cooperation in multi-agent reinforcement learning (MARL) based upon any nonlinear function of the team's long-term state-action occupancy measure, i.e., a \emph{general utility}. This subsumes the cumulative return but also allows one to incorporate risk-sensitivity, exploration, and priors. % We derive the {\bf D}ecentralized {\bf S}hadow Reward {\bf A}ctor-{\bf C}ritic (DSAC) in which agents alternate between policy evaluation (critic), weighted averaging with neighbors (information mixing), and local gradient updates for their policy parameters (actor). DSAC augments the classic critic step by requiring agents to (i) estimate their local occupancy measure in order to (ii) estimate the derivative of the local utility with respect to their occupancy measure, i.e., the ``shadow reward". DSAC converges to $\epsilon$-stationarity in $\mathcal{O}(1/\epsilon^{2.5})$ (Theorem \ref{theorem:final}) or faster $\mathcal{O}(1/\epsilon^{2})$ (Corollary \ref{corollary:communication}) steps with high probability, depending on the amount of communications. We further establish the non-existence of spurious stationary points for this problem, that is, DSAC finds the globally optimal policy (Corollary \ref{corollary:global}). Experiments demonstrate the merits of goals beyond the cumulative return in cooperative MARL.

poster session 1 @ blue 1, poster session 8 @ blue 1, poster session 1, poster session 8

Offline Reinforcement Learning (RL) aims at learning an optimal control from a fixed dataset, without interactions with the system. An agent in this setting should avoid selecting actions whose consequences cannot be predicted from the data. This is the converse of exploration in RL, which favors such actions. We thus take inspiration from the literature on bonus-based exploration to design a new offline RL agent. The core idea is to subtract a prediction-based exploration bonus from the reward, instead of adding it for exploration. This allows the policy to stay close to the support of the dataset and practically extends some previous pessimism-based offline RL methods to a deep learning setting with arbitrary bonuses. We also connect this approach to a more common regularization of the learned policy towards the data. Instantiated with a bonus based on the prediction error of a variational autoencoder, we show that our simple agent is competitive with the state of the art on a set of continuous control locomotion and manipulation tasks.

poster session 1 @ blue 1, poster session 8 @ blue 1, poster session 1, poster session 8

We present a novel approach to hierarchical reinforcement learning for linearly-solvable Markov decision processes. Our approach assumes that the state space is partitioned, and defines subtasks for moving between the partitions. We represent value functions on several levels of abstraction, and use the compositionality of subtasks to estimate the optimal values of the states in each partition. The policy is implicitly defined on these optimal value estimates, rather than being decomposed among the subtasks. As a consequence, our approach can learn the globally optimal policy, and does not suffer from non-stationarities induced by high-level decisions. If several partitions have equivalent dynamics, the subtasks of those partitions can be shared. We show that our approach is significantly more sample efficient than that of a flat learner and similar hierarchical approaches when the set of boundary states is smaller than the entire state space.

poster session 1 @ blue 1, poster session 8 @ blue 1, poster session 1, poster session 8

Robust adversarial reinforcement learning is an effective method to train agents to manage uncertain disturbance and modeling errors in real environments. However, for systems that are sensitive to disturbances or those that are difficult to stabilize, it is easier to learn a powerful adversary than establish a stable control policy. An improper strong adversary can destabilize the system, introduce biases in the sampling process, make the learning process unstable, and even reduce the robustness of the policy. In this study, we consider the problem of ensuring system stability during training in the adversarial reinforcement learning architecture. The dissipative principle of robust $H_\infty$ control is extended to the Markov Decision Process, and robust stability constraints are obtained based on $L_2$ gain performance in the reinforcement learning system. Thus, we propose a dissipation-inequation-constraint-based adversarial reinforcement learning architecture. This architecture ensures the stability of the system during training by imposing constraints on the normal and adversarial agents. Theoretically, this architecture can be applied to a large family of deep reinforcement learning algorithms. Results of experiments in MuJoCo and GymFc environments show that our architecture effectively improves the robustness of the controller against environmental changes and adapts to more powerful adversaries. Results of the flight experiments on a real quadcopter indicate that our method can directly deploy the policy trained in the simulation environment to the real environment, and our controller outperforms the PID controller based on hardware-in-the-loop. Both our theoretical and empirical results provide new and critical outlooks on the adversarial reinforcement learning architecture from a rigorous robust control perspective.

poster session 1 @ blue 1, poster session 8 @ blue 1, poster session 1, poster session 8

The shortcomings of maximum likelihood estimation in the context of model-based reinforcement learning have been highlighted by an increasing number of papers. When the model class is misspecified or has a limited representational capacity, model parameters with high likelihood might not necessarily result in high performance of the agent on a downstream control task. To alleviate this problem, we propose an end-to-end approach for model learning which directly optimizes the expected returns using implicit differentiation. We treat a value function that satisfies the Bellman optimality operator induced by the model as an implicit function of model parameters and show how to differentiate the function. We provide theoretical and empirical evidence highlighting the benefits of our approach in the model misspecification regime compared to likelihood-based methods.

poster session 1 @ blue 1, poster session 8 @ blue 1, poster session 1, poster session 8

Estimating value functions is a core component of reinforcement learning algorithms. Temporal difference (TD) learning algorithms use bootstrapping, i.e. they update the value function toward a learning target using value estimates at subsequent time-steps. Alternatively, the value function can be updated toward a learning target constructed by separately predicting successor features (SF)--a policy-dependent model--and linearly combining them with instantaneous rewards. We focus on bootstrapping targets used when estimating value functions, and propose a new backup target, the $\eta$-return mixture, which implicitly combines value-predictive knowledge (used by TD methods) with (successor) feature-predictive knowledge--with a parameter $\eta$ capturing how much to rely on each. We illustrate that incorporating predictive knowledge through an $\eta\gamma$-discounted SF model makes more efficient use of sampled experience, compared to either extreme, i.e. bootstrapping entirely on the value function estimate, or bootstrapping on the product of separately estimated successor features and instantaneous reward models. We empirically show this approach leads to faster policy evaluation and better control performance, for tabular and nonlinear function approximations, indicating scalability and generality.

poster session 1 @ blue 1, poster session 8 @ blue 1, poster session 1, poster session 8

Model-free reinforcement learning (RL) algorithms are known to be memory and computation-efficient compared with model-based approaches. This paper presents a model-free RL algorithm for infinite-horizon average-reward Constrained Markov Decision Processes (CMDPs), which achieves sub-linear regret and zero constraint violation. To the best of our knowledge, this is the first {\em model-free} algorithm for general CMDPs in the infinite-horizon average-reward setting with provable guarantees.

poster session 1 @ blue 1, poster session 8 @ blue 1, poster session 1, poster session 8

We study the problem of safe offline reinforcement learning (RL), the goal is to learn a policy that maximizes long-term reward while satisfying safety constraints given only offline data, without further interaction with the environment. This problem is more appealing for real world RL applications, in which data collection is costly or dangerous. Enforcing constraint satisfaction is non-trivial, especially in offline settings, as there is a potential large discrepancy between the policy distribution and the data distribution, causing errors in estimating the value of safety constraints. We show that na\"ive approaches that combine techniques from safe RL and offline RL can only learn sub-optimal solutions. We thus develop a simple yet effective algorithm, Constraints Penalized Q-Learning (CPQ), to solve the problem. Our method admits the use of data generated by mixed behavior policies. We present a theoretical analysis and demonstrate empirically that our approach can learn robustly across a variety of benchmark control tasks, outperforming several baselines.

poster session 1 @ blue 1, poster session 8 @ blue 1, poster session 1, poster session 8

Inspired by the extensive success of deep learning, graph neural networks (GNNs) have been proposed to learn expressive node representations and demonstrated promising performance in various graph learning tasks. However, existing endeavors predominately focus on the conventional semi-supervised setting where relatively abundant gold-labeled nodes are provided. While it is often impractical due to the fact that data labeling is unbearably laborious and requires intensive domain knowledge, especially when considering the heterogeneity of graph-structured data. Under the few-shot semi-supervised setting, the performance of most of the existing GNNs is inevitably undermined by the overfitting and oversmoothing issues, largely owing to the shortage of labeled data. In this paper, we propose a decoupled network architecture equipped with a novel meta-learning algorithm to solve this problem. In essence, our framework Meta-PN infers high-quality pseudo labels on unlabeled nodes via a meta-learned label propagation strategy, which effectively augments the scarce labeled data while enabling large receptive fields during training. Extensive experiments demonstrate that our approach offers easy and substantial performance gains compared to existing techniques on various benchmark datasets.

poster session 1 @ blue 1, poster session 8 @ blue 1, oral session 8 @ blue 1, poster session 1, poster session 8, oral session 8

Unsupervised graph-level representation learning plays a crucial role in a variety of tasks such as molecular property prediction and community analysis, especially when data annotation is expensive. Currently, most of the best-performing graph embedding methods are based on Infomax principle. The performance of these methods highly depends on the selection of negative samples and hurt the performance, if the samples were not carefully selected. Inter-graph similarity-based methods also suffer if the selected set of graphs for similarity matching is low in quality. To address this, we focus only on utilizing the current input graph for embedding learning. We are motivated by an observation from real-world graph generation processes where the graphs are formed based on one or more global factors which are common to all elements of the graph (e.g., topic of a discussion thread, solubility level of a molecule). We hypothesize extracting these common factors could be highly beneficial. Hence, this work proposes a new principle for unsupervised graph representation learning: Graph-wise Common latent Factor EXtraction (GCFX). We further propose a deep model for GCFX, deepGCFX, based on the idea of reversing the above-mentioned graph generation process which could explicitly extract common latent factors from an input graph and achieve improved results on downstream tasks to the current state-of-the-art. Through extensive experiments and analysis, we demonstrate that, while extracting common latent factors is beneficial for graph-level tasks to alleviate distractions caused by local variations of individual nodes or local neighbourhoods, it also benefits node-level tasks by enabling long-range node dependencies, especially for disassortative graphs.

poster session 1 @ blue 1, poster session 8 @ blue 1, oral session 8 @ blue 1, poster session 1, poster session 8, oral session 8

Hierarchical reinforcement learning has focused on discovering temporally extended actions, such as options, that can provide benefits in problems requiring extensive exploration. One promising approach that learns these options end-to-end is the option-critic (OC) framework. We examine and show in this paper that OC does not decompose a problem into simpler sub-problems, but instead increases the size of the search over policy space with each option considering the entire state space during learning. This issue can result in practical limitations of this method, including sample inefficient learning. To address this problem, we introduce Context-Specific Representation Abstraction for Deep Option Learning (CRADOL), a new framework that considers both temporal abstraction and context-specific representation abstraction to effectively reduce the size of the search over policy space. Specifically, our method learns a factored belief state representation that enables each option to learn a policy over only a subsection of the state space. We test our method against hierarchical, non-hierarchical, and modular recurrent neural network baselines, demonstrating significant sample efficiency improvements in challenging partially observable environments.

poster session 1 @ blue 1, poster session 8 @ blue 1, poster session 1, poster session 8

Reinforcement learning (RL) has demonstrated its superiority in solving sequential decision-making problems. However, heavy dependence on immediate reward feedback impedes the wide application of RL. On the other hand, imitation learning (IL) tackles RL without relying on environmental supervision by leveraging external demonstrations. In practice, however, collecting sufficient expert demonstrations can be prohibitively expensive, yet the quality of demonstrations typically limits the performance of the learning policy. To address a practical scenario, in this work, we propose Self-Adaptive Imitation Learning (SAIL), which, provided with a few demonstrations from a sub-optimal teacher, can perform well in RL tasks with extremely delayed rewards, where the only reward feedback is trajectory-wise ranking. SAIL bridges the advantages of IL and RL by interactively exploiting the demonstrations to catch up with the teacher and exploring the environment to yield demonstrations that surpass the teacher. Extensive empirical results show that not only does SAIL significantly improve the sample efficiency, but it also leads to higher asymptotic performance across different continuous control tasks, compared with the state-of-the-art.

poster session 1 @ blue 1, poster session 8 @ blue 1, poster session 1, poster session 8