ICML 2023 论文列表

Model-based reinforcement learning (RL) often achieves higher sample efficiency in practice than model-free RL by learning a dynamics model to generate samples for policy learning. Previous works learn a dynamics model that fits under the empirical state-action visitation distribution for all historical policies, i.e., the sample replay buffer. However, in this paper, we observe that fitting the dynamics model under the distribution for *all historical policies* does not necessarily benefit model prediction for the *current policy* since the policy in use is constantly evolving over time. The evolving policy during training will cause state-action visitation distribution shifts. We theoretically analyze how this distribution shift over historical policies affects the model learning and model rollouts. We then propose a novel dynamics model learning method, named *Policy-adapted Dynamics Model Learning (PDML)*. PDML dynamically adjusts the historical policy mixture distribution to ensure the learned model can continually adapt to the state-action visitation distribution of the evolving policy. Experiments on a range of continuous control environments in MuJoCo show that PDML achieves significant improvement in sample efficiency and higher asymptotic performance combined with the state-of-the-art model-based RL methods.

Algorithms for safely improving policies are important to deploy reinforcement learning approaches in real-world scenarios. In this work, we propose an algorithm, called MCTS-SPIBB, that computes safe policy improvement online using a Monte Carlo Tree Search based strategy. We theoretically prove that the policy generated by MCTS-SPIBB converges, as the number of simulations grows, to the optimal safely improved policy generated by Safe Policy Improvement with Baseline Bootstrapping (SPIBB), a popular algorithm based on policy iteration. Moreover, our empirical analysis performed on three standard benchmark domains shows that MCTS-SPIBB scales to significantly larger problems than SPIBB because it computes the policy online and locally, i.e., only in the states actually visited by the agent.

Trustworthy policy learning has significant importance in making reliable and harmless treatment decisions for individuals. Previous policy learning approaches aim at the well-being of subgroups by maximizing the utility function (e.g., conditional average causal effects, post-view click-through&conversion rate in recommendations), however, individual-level counterfactual no-harm criterion has rarely been discussed. In this paper, we first formalize the counterfactual no-harm criterion for policy learning from a principal stratification perspective. Next, we propose a novel upper bound for the fraction negatively affected by the policy and show the consistency and asymptotic normality of the estimator. Based on the estimators for the policy utility and harm upper bounds, we further propose a policy learning approach that satisfies the counterfactual no-harm criterion, and prove its consistency to the optimal policy reward for parametric and non-parametric policy classes, respectively. Extensive experiments are conducted to show the effectiveness of the proposed policy learning approach for satisfying the counterfactual no-harm criterion.

We study off-policy evaluation (OPE) of contextual bandit policies for large discrete action spaces where conventional importance-weighting approaches suffer from excessive variance. To circumvent this variance issue, we propose a new estimator, called *OffCEM*, that is based on the *conjunct effect model* (CEM), a novel decomposition of the causal effect into a cluster effect and a residual effect. OffCEM applies importance weighting only to action clusters and addresses the residual causal effect through model-based reward estimation. We show that the proposed estimator is unbiased under a new assumption, called *local correctness*, which only requires that the residual-effect model preserves the relative expected reward differences of the actions within each cluster. To best leverage the CEM and local correctness, we also propose a new two-step procedure for performing model-based estimation that minimizes bias in the first step and variance in the second step. We find that the resulting OffCEM estimator substantially improves bias and variance compared to a range of conventional estimators. Experiments demonstrate that OffCEM provides substantial improvements in OPE especially in the presence of many actions.

In model-based reinforcement learning (MBRL), policy gradients can be estimated either by derivative-free RL methods, such as likelihood ratio gradients (LR), or by backpropagating through a differentiable model via reparameterization gradients (RP). Instead of using one or the other, the Total Propagation (TP) algorithm in prior work showed that a combination of LR and RP estimators averaged using inverse variance weighting (IVW) can achieve orders of magnitude improvement over either method. However, IVW-based composite estimators have not yet been applied in modern RL tasks, as it is unclear if they can be implemented scalably. We propose a scalable method, Total Propagation X (TPX) that improves over TP by changing the node used for IVW, and employing coordinate wise weighting. We demonstrate the scalability of TPX by applying it to the state of the art visual MBRL algorithm Dreamer. The experiments showed that Dreamer fails with long simulation horizons, while our TPX works reliably for only a fraction of additional computation. One key advantage of TPX is its ease of implementation, which will enable experimenting with IVW on many tasks beyond MBRL.

Exploration remains a key challenge in deep reinforcement learning (RL). Optimism in the face of uncertainty is a well-known heuristic with theoretical guarantees in the tabular setting, but how best to translate the principle to deep reinforcement learning, which involves online stochastic gradients and deep network function approximators, is not fully understood. In this paper we propose a new, differentiable optimistic objective that when optimized yields a policy that provably explores efficiently, with guarantees even under function approximation. Our new objective is a zero-sum two-player game derived from endowing the agent with an epistemic-risk-seeking utility function, which converts uncertainty into value and encourages the agent to explore uncertain states. We show that the solution to this game minimizes an upper bound on the regret, with the 'players' each attempting to minimize one component of a particular regret decomposition. We derive a new model-free algorithm which we call 'epistemic-risk-seeking actor-critic' (ERSAC), which is simply an application of simultaneous stochastic gradient ascent-descent to the game. Finally, we discuss a recipe for incorporating off-policy data and show that combining the risk-seeking objective with replay data yields a double benefit in terms of statistical efficiency. We conclude with some results showing good performance of a deep RL agent using the technique on the challenging 'DeepSea' environment, showing significant performance improvements even over other efficient exploration techniques, as well as improved performance on the Atari benchmark.

A *treatment policy* defines when and what treatments are applied to affect some outcome of interest. Data-driven decision-making requires the ability to predict *what happens if a policy is changed*. Existing methods that predict how the outcome evolves under different scenarios assume that the tentative sequences of future treatments are fixed in advance, while in practice the treatments are determined stochastically by a policy and may depend, for example, on the efficiency of previous treatments. Therefore, the current methods are not applicable if the treatment policy is unknown or a counterfactual analysis is needed. To handle these limitations, we model the treatments and outcomes jointly in continuous time, by combining Gaussian processes and point processes. Our model enables the estimation of a treatment policy from observational sequences of treatments and outcomes, and it can predict the interventional and counterfactual progression of the outcome *after an intervention on the treatment policy* (in contrast with the causal effect of a single treatment). We show with real-world and semi-synthetic data on blood glucose progression that our method can answer causal queries more accurately than existing alternatives.

We consider the task of evaluating policies of algorithmic resource allocation through randomized controlled trials (RCTs). Such policies are tasked with optimizing the utilization of limited intervention resources, with the goal of maximizing the benefits derived. Evaluation of such allocation policies through RCTs proves difficult, notwithstanding the scale of the trial, because the individuals’ outcomes are inextricably interlinked through resource constraints controlling the policy decisions. Our key contribution is to present a new estimator leveraging our proposed novel concept, that involves retrospective reshuffling of participants across experimental arms at the end of an RCT. We identify conditions under which such reassignments are permissible and can be leveraged to construct counterfactual trials, whose outcomes can be accurately ascertained, for free. We prove theoretically that such an estimator is more accurate than common estimators based on sample means -- we show that it returns an unbiased estimate and simultaneously reduces variance. We demonstrate the value of our approach through empirical experiments on synthetic, semisynthetic as well as real case study data and show improved estimation accuracy across the board.

Behavior constrained policy optimization has been demonstrated to be a successful paradigm for tackling Offline Reinforcement Learning. By exploiting historical transitions, a policy is trained to maximize a learned value function while constrained by the behavior policy to avoid a significant distributional shift. In this paper, we propose our closed-form policy improvement operators. We make a novel observation that the behavior constraint naturally motivates the use of first-order Taylor approximation, leading to a linear approximation of the policy objective. Additionally, as practical datasets are usually collected by heterogeneous policies, we model the behavior policies as a Gaussian Mixture and overcome the induced optimization difficulties by leveraging the LogSumExp's lower bound and Jensen's Inequality, giving rise to a closed-form policy improvement operator. We instantiate both one-step and iterative offline RL algorithms with our novel policy improvement operators and empirically demonstrate their effectiveness over state-of-the-art algorithms on the standard D4RL benchmark. Our code is available at https://cfpi-icml23.github.io/.

It is often very challenging to manually design reward functions for complex, real-world tasks. To solve this, one can instead use reward learning to infer a reward function from data. However, there are often multiple reward functions that fit the data equally well, even in the infinite-data limit. This means that the reward function is only partially identifiable. In this work, we formally characterise the partial identifiability of the reward function given several popular reward learning data sources, including expert demonstrations and trajectory comparisons. We also analyse the impact of this partial identifiability for several downstream tasks, such as policy optimisation. We unify our results in a framework for comparing data sources and downstream tasks by their invariances, with implications for the design and selection of data sources for reward learning.

We study a class of reinforcement learning problems where the reward signals for policy learning are generated by a discriminator that is dependent on and jointly optimized with the policy. This interdependence between the policy and the discriminator leads to an unstable learning process because reward signals from an immature discriminator are noisy and impede policy learning, and conversely, an under-optimized policy impedes discriminator learning. We call this learning setting $\textit{Internally Rewarded Reinforcement Learning}$ (IRRL) as the reward is not provided directly by the environment but $\textit{internally}$ by the discriminator. In this paper, we formally formulate IRRL and present a class of problems that belong to IRRL. We theoretically derive and empirically analyze the effect of the reward function in IRRL and based on these analyses propose the clipped linear reward function. Experimental results show that the proposed reward function can consistently stabilize the training process by reducing the impact of reward noise, which leads to faster convergence and higher performance compared with baselines in diverse tasks.

Risk-sensitive reinforcement learning (RL) has become a popular tool to control the risk of uncertain outcomes and ensure reliable performance in various sequential decision-making problems. While policy gradient methods have been developed for risk-sensitive RL, it remains unclear if these methods enjoy the same global convergence guarantees as in the risk-neutral case. In this paper, we consider a class of dynamic time-consistent risk measures, called Expected Conditional Risk Measures (ECRMs), and derive policy gradient updates for ECRM-based objective functions. Under both constrained direct parameterization and unconstrained softmax parameterization, we provide global convergence and iteration complexities of the corresponding risk-averse policy gradient algorithms. We further test risk-averse variants of REINFORCE and actor-critic algorithms to demonstrate the efficacy of our method and the importance of risk control.

Recently, the impressive empirical success of policy gradient (PG) methods has catalyzed the development of their theoretical foundations. Despite the huge efforts directed at the design of efficient stochastic PG-type algorithms, the understanding of their convergence to a globally optimal policy is still limited. In this work, we develop improved global convergence guarantees for a general class of Fisher-non-degenerate parameterized policies which allows to address the case of continuous state action spaces. First, we propose a Normalized Policy Gradient method with Implicit Gradient Transport (N-PG-IGT) and derive a $\tilde{\mathcal{O}}(\varepsilon^{-2.5})$ sample complexity of this method for finding a global $\varepsilon$-optimal policy. Improving over the previously known $\tilde{\mathcal{O}}(\varepsilon^{-3})$ complexity, this algorithm does not require the use of importance sampling or second-order information and samples only one trajectory per iteration. Second, we further improve this complexity to $\tilde{ \mathcal{\mathcal{O}} }(\varepsilon^{-2})$ by considering a Hessian-Aided Recursive Policy Gradient ((N)-HARPG) algorithm enhanced with a correction based on a Hessian-vector product. Interestingly, both algorithms are $(i)$ simple and easy to implement: single-loop, do not require large batches of trajectories and sample at most two trajectories per iteration; $(ii)$ computationally and memory efficient: they do not require expensive subroutines at each iteration and can be implemented with memory linear in the dimension of parameters.

Policy Optimization (PO) is one of the most popular methods in Reinforcement Learning (RL). Thus, theoretical guarantees for PO algorithms have become especially important to the RL community. In this paper, we study PO in adversarial MDPs with a challenge that arises in almost every real-world application -- *delayed bandit feedback*. We give the first near-optimal regret bounds for PO in tabular MDPs, and may even surpass state-of-the-art (which uses less efficient methods). Our novel Delay-Adapted PO (DAPO) is easy to implement and to generalize, allowing us to extend our algorithm to: (i) infinite state space under the assumption of linear $Q$-function, proving the first regret bounds for delayed feedback with function approximation. (ii) deep RL, demonstrating its effectiveness in experiments on MuJoCo domains.

Policy optimization methods are popular reinforcement learning algorithms in practice and recent works have build theoretical foundation for them by proving $\sqrt{T}$ regret bounds even when the losses are adversarial. Such bounds are tight in the worst case but often overly pessimistic. In this work, we show that by carefully designing the regularizer, bonus terms, and learning rates, one can achieve a more favorable $\text{polylog}(T)$ regret bound when the losses are stochastic, without sacrificing the worst-case guarantee in the adversarial regime. Specifically, we show the first best of both worlds guarantee for policy optimization in tabular MDPs by leveraging either a Tsallis entropy or a Shannon entropy regularizer. Then we show that under known transitions, we can further obtain a first-order regret bound in the adversarial regime by leveraging the log barrier regularizer.

Off-policy evaluation (OPE) aims to estimate the return of a target policy using some pre-collected observational data generated by a potentially different behavior policy. In many cases, there exist unmeasured variables that confound the action-reward or action-next-state relationships, rendering many existing OPE approaches ineffective. This paper develops an instrumental variable (IV)-based method for consistent OPE in confounded sequential decision making. Similar to single-stage decision making, we show that IV enables us to correctly identify the target policy's value in infinite horizon settings as well. Furthermore, we propose a number of policy value estimators and illustrate their effectiveness through extensive simulations and real data analysis from a world-leading short-video platform.

Masked image modeling, an emerging self-supervised pre-training method, has shown impressive success across numerous downstream vision tasks with Vision transformers. Its underlying idea is simple: a portion of the input image is masked out and then reconstructed via a pre-text task. However, the working principle behind MIM is not well explained, and previous studies insist that MIM primarily works for the Transformer family but is incompatible with CNNs. In this work, we observe that MIM essentially teaches the model to learn better middle-order interactions among patches for more generalized feature extraction. We then propose an Architecture-Agnostic Masked Image Modeling framework (A$^2$MIM), which is compatible with both Transformers and CNNs in a unified way. Extensive experiments on popular benchmarks show that A$^2$MIM learns better representations without explicit design and endows the backbone model with the stronger capability to transfer to various downstream tasks.

In model-free reinforcement learning, recent methods based on a phasic policy gradient (PPG) framework have shown impressive improvements in sample efficiency and zero-shot generalization on the challenging Procgen benchmark. In PPG, two design choices are believed to be the key contributing factors to its superior performance over PPO: the high level of value sample reuse and the low frequency of feature distillation. However, through an extensive empirical study, we unveil that policy regularization and data diversity are what actually matters. In particular, we can achieve the same level of performance with low value sample reuse and frequent feature distillation, as long as the policy regularization strength and data diversity are preserved. In addition, we can maintain the high performance of PPG while reducing the computational cost to a similar level as PPO. Our comprehensive study covers all 16 Procgen games in both sample efficiency and generalization setups. We hope it can advance the understanding of PPG and provide insights for future works.

The average reward criterion is relatively less studied as most existing works in the Reinforcement Learning literature consider the discounted reward criterion. There are few recent works that present on-policy average reward actor-critic algorithms, but average reward off-policy actor-critic is relatively less explored. In this work, we present both on-policy and off-policy deterministic policy gradient theorems for the average reward performance criterion. Using these theorems, we also present an Average Reward Off-Policy Deep Deterministic Policy Gradient (ARO-DDPG) Algorithm. We first show asymptotic convergence analysis using the ODE-based method. Subsequently, we provide a finite time analysis of the resulting stochastic approximation scheme with linear function approximator and obtain an $\epsilon$-optimal stationary policy with a sample complexity of $\Omega(\epsilon^{-2.5})$. We compare the average reward performance of our proposed ARO-DDPG algorithm and observe better empirical performance compared to state-of-the-art on-policy average reward actor-critic algorithms over MuJoCo-based environments.