In real-world decision-making tasks, it is critical for reinforcement learning (RL) methods to be both stable and robust. Maximum entropy RL methods typically generate a robust policy with entropy augmented reward. While incorporating entropy into the reward offers the benefit of exploration, it presents limited universal applicability and persistent convergence difficulties, such as suboptimal policy stabilization and unstable $Q$ value update. From optimization, we define these two issues as tremulous policy and spiky Q-function, investigating their underlying causes and relationships. Analysis with this, the maximum entropy principle leads to a spiky $Q$ -function update, which ultimately results in a tremulous policy. We thus introduce a beta-symmetric Kullback-Leibler (KL) divergence objective to mitigate such issues under the maximum entropy framework. With this objective function, the tremulous nature of the policy could be controlled with a large beta value. The spiky $Q$ -function could be avoided by annealing the entropy in the target $Q$ value, as the beta-symmetric KL divergence is an upper bound of the original reverse KL divergence. Theoretically, we prove that minimizing our new objective function results in a new policy that presents an improvement in the $Q$ value. Guaranteed by these results, we ultimately derive the optimal policy by iteratively updating the $Q$ value and policy, and we call this method max-entropy stable optimization (MeSO). Experimental results on the Mujoco and Roboschool platforms demonstrate that our algorithm maintains stability while offering better flexibility and overall performance.

The popular Proximal Policy Optimization (PPO) algorithm approximates the solution in a clipped policy space. Does there exist better policies outside of this space? By using a novel surrogate objective that employs the sigmoid function (which provides an interesting way of exploration), we found that the answer is "YES", and the better policies are in fact located very far from the clipped space. We show that PPO is insufficient in "off-policyness", according to an off-policy metric called DEON. Our algorithm explores in a much larger policy space than PPO, and it maximizes the Conservative Policy Iteration (CPI) objective better than PPO during training. To the best of our knowledge, all current PPO methods have the clipping operation and optimize in the clipped policy space. Our method is the first of this kind, which advances the understanding of CPI optimization and policy gradient methods. Code is available at https://github.com/raincchio/P3O.

Prioritized Experience Replay (ER) has been empirically shown to improve sample efficiency across many domains and attracted great attention; however, there is little theoretical understanding of why such prioritized sampling helps and its limitations. In this work, we take a deep look at the prioritized ER. In a supervised learning setting, we show the equivalence between the error-based prioritized sampling method for minimizing mean squared error and the uniform sampling for cubic power loss. We then provide theoretical insight into why error-based prioritized sampling improves convergence rate upon uniform sampling when minimizing mean squared error during early learning. Based on the insight, we further point out two limitations of the prioritized ER method: 1) outdated priorities and 2) insufficient coverage of the sample space. To mitigate the limitations, we propose our model-based stochastic gradient Langevin dynamics sampling method. We show that our method does provide states distributed close to an ideal prioritized sampling distribution estimated by the brute-force method, which does not suffer from the two limitations. We conduct experiments on both discrete and continuous control problems to show our approach’s efficacy and examine the practical implication of our method in an autonomous driving application.

The deadly triad refers to the instability of a reinforcement learning algorithm when it employs off-policy learning, function approximation, and bootstrapping simultaneously. In this paper, we investigate the target network as a tool for breaking the deadly triad, providing theoretical support for the conventional wisdom that a target network stabilizes training. We first propose and analyze a novel target network update rule which augments the commonly used Polyak-averaging style update with two projections. We then apply the target network and ridge regularization in several divergent algorithms and show their convergence to regularized TD fixed points. Those algorithms are off-policy with linear function approximation and bootstrapping, spanning both policy evaluation and control, as well as both discounted and average-reward settings. In particular, we provide the first convergent linear $Q$-learning algorithms under nonrestrictive and changing behavior policies without bi-level optimization.

We present the first provably convergent two-timescale off-policy actor-critic algorithm (COF-PAC) with function approximation. Key to COF-PAC is the introduction of a new critic, the emphasis critic, which is trained via Gradient Emphasis Learning (GEM), a novel combination of the key ideas of Gradient Temporal Difference Learning and Emphatic Temporal Difference Learning. With the help of the emphasis critic and the canonical value function critic, we show convergence for COF-PAC, where the critics are linear and the actor can be nonlinear.

We consider the problem of mapless collision-avoidance navigation where humans are present using 2D laser scans. Our proposed method uses ego-safety to measure collision from the robot's perspective and social-safety to measure the impact of robot's actions on surrounding pedestrians. Specifically, the social-safety part predicts the intrusion impact of the robot's action into the interaction area with surrounding humans. We train the policy using reinforcement learning on a simple simulator and directly evaluate the learned policy in Gazebo and real robot tests. Experiments show the learned policy smoothly transferred to different scenarios without any fine tuning. We observe that our method demonstrates time-efficient path planning behavior with high success rate in the mapless navigation task. Furthermore, we test our method in a navigation task among dynamic crowds, considering both low and high volume traffic. Our learned policy demonstrates cooperative behavior that actively drives our robot into traffic flows while showing respect to nearby pedestrians. Evaluation videos are at https://sites.google.com/view/ssw-batman.

Significant progress has been made recently in developing few-shot object segmentation methods. Learning is shown to be successful in few-shot segmentation settings, using pixel-level, scribbles and bounding box supervision. This paper takes another approach, i.e., only requiring image-level label for few-shot object segmentation. We propose a novel multi-modal interaction module for few-shot object segmentation that utilizes a co-attention mechanism using both visual and word embedding. Our model using image-level labels achieves 4.8% improvement over previously proposed image-level few-shot object segmentation. It also outperforms state-of-the-art methods that use weak bounding box supervision on PASCAL-5^i. Our results show that few-shot segmentation benefits from utilizing word embeddings, and that we are able to perform few-shot segmentation using stacked joint visual semantic processing with weak image-level labels. We further propose a novel setup, Temporal Object Segmentation for Few-shot Learning (TOSFL) for videos. TOSFL can be used on a variety of public video data such as Youtube-VOS, as demonstrated in both instance-level and category-level TOSFL experiments.

In this paper, we propose an actor ensemble algorithm, named ACE, for continuous control with a deterministic policy in reinforcement learning. In ACE, we use actor ensemble (i.e., multiple actors) to search the global maxima of the critic. Besides the ensemble perspective, we also formulate ACE in the option framework by extending the option-critic architecture with deterministic intra-option policies, revealing a relationship between ensemble and options. Furthermore, we perform a look-ahead tree search with those actors and a learned value prediction model, resulting in a refined value estimation. We demonstrate a significant performance boost of ACE over DDPG and its variants in challenging physical robot simulators.

In this paper, we propose the Quantile Option Architecture (QUOTA) for exploration based on recent advances in distributional reinforcement learning (RL). In QUOTA, decision making is based on quantiles of a value distribution, not only the mean. QUOTA provides a new dimension for exploration via making use of both optimism and pessimism of a value distribution. We demonstrate the performance advantage of QUOTA in both challenging video games and physical robot simulators.

Many problems in science and engineering can be reduced to the recovery of an unknown large matrix from a small number of random linear measurements. Matrix factorization arguably is the most popular approach for low-rank matrix recovery. Many methods have been proposed using different loss functions, for example the most widely used L2loss, more robust choices such as L1and Huber loss, quantile and expectile loss for skewed data. All of them can be unified into the framework of M-estimation. In this paper, we present a general framework of low-rank matrix factorization based on M-estimation in statistics. The framework mainly involves two steps: firstly we apply Nesterov's smoothing technique to obtain an optimal smooth approximation for non-smooth loss function, such as L1and quantile loss; secondly we exploit an alternative updating scheme along with Nesterov's momentum method at each step to minimize the smoothed loss function. Strong theoretical convergence guarantee has been developed for the general framework, and extensive numerical experiments have been conducted to illustrate the performance of proposed algorithm.

In distributional reinforcement learning (RL), the estimated distribution of value functions model both the parametric and intrinsic uncertainties. We propose a novel and efficient exploration method for deep RL that has two components. The first is a decaying schedule to suppress the intrinsic uncertainty. The second is an exploration bonus calculated from the upper quantiles of the learned distribution. In Atari 2600 games, our method achieves 483 % average gain across 49 games in cumulative rewards over QR-DQN. We also compared our algorithm with QR-DQN in a challenging 3D driving simulator (CARLA). Results show that our algorithm achieves nearoptimal safety rewards twice faster than QRDQN.

Dyna is an architecture for model based reinforcement learning (RL), where simulated experience from a model is used to update policies or value functions. A key component of Dyna is search control, the mechanism to generate the state and action from which the agent queries the model, which remains largely unexplored. In this work, we propose to generate such states by using the trajectory obtained from Hill Climbing (HC) the current estimate of the value function. This has the effect of propagating value from high value regions and of preemptively updating value estimates of the regions that the agent is likely to visit next. We derive a noisy projected natural gradient algorithm for hill climbing, and highlight a connection to Langevin dynamics. We provide an empirical demonstration on four classical domains that our algorithm, HC Dyna, can obtain significant sample efficiency improvements. We study the properties of different sampling distributions for search control, and find that there appears to be a benefit specifically from using the samples generated by climbing on current value estimates from low value to high value region.

In this paper we introduce the concept of pseudo-MDPs to develop abstractions. Pseudo-MDPs relax the requirement that the transition kernel has to be a probability kernel. We show that the new framework captures many existing abstractions. We also introduce the concept of factored linear action models; a special case. Again, the relation of factored linear action models and existing works are discussed. We use the general framework to develop a theory for bounding the suboptimality of policies derived from pseudo-MDPs. Specializing the framework, we recover existing results. We give a leastsquares approach and a constrained optimization approach of learning the factored linear model as well as efficient computation methods. We demonstrate that the constrained optimization approach gives better performance than the least-squares approach with normalization.

In this paper we consider the problem of finding a good policy given some batch data.We propose a new approach, LAM-API, that first builds a so-called linear action model (LAM) from the data and then uses the learned model and the collected data in approximate policy iteration (API) to find a good policy.A natural choice for the policy evaluation step in this algorithm is to use least-squares temporal difference (LSTD) learning algorithm.Empirical results on three benchmark problems show that this particular instance of LAM-API performs competitively as compared with LSPI, both from the point of view of data and computational efficiency.

We extend Dyna planning architecture for policy evaluation and control in two significant aspects. First, we introduce a multi-step Dyna planning that projects the simulated state/feature many steps into the future. Our multi-step Dyna is based on a multi-step model, which we call the {\em $\lambda$-model}. The $\lambda$-model interpolates between the one-step model and an infinite-step model, and can be learned efficiently online. Second, we use for Dyna control a dynamic multi-step model that is able to predict the results of a sequence of greedy actions and track the optimal policy in the long run. Experimental results show that Dyna using the multi-step model evaluates a policy faster than using single-step models; Dyna control algorithms using the dynamic tracking model are much faster than model-free algorithms; further, multi-step Dyna control algorithms enable the policy and value function to converge much faster to their optima than single-step Dyna algorithms.

This paper extends many of the recent popular policy evaluation algorithms to a generalized framework that includes least-squares temporal difference (LSTD) learning, leastsquares policy evaluation (LSPE) and a variant of incremental LSTD (iLSTD). The basis of this extension is a preconditioning technique that solves a stochastic model equation. This paper also studies three significant issues of the new framework: it presents a new rule of step-size that can be computed online, provides an iterative way to apply preconditioning, and reduces the complexity of related algorithms to near that of temporal difference (TD) learning. 1.