The dominant framework for off-policy multi-goal reinforcement learning involves estimating goal conditioned Q-value function. When learning to achieve multiple goals, data efficiency is intimately connected with the generalization of the Q-function to new goals. The de-facto paradigm is to approximate Q(s, a, g) using monolithic neural networks. To improve the generalization of the Q-function, we propose a bilinear decomposition that represents the Q-value via a low-rank approximation in the form of a dot product between two vector fields. The first vector field, f(s, a), captures the environment's local dynamics at the state s; whereas the second component, {\phi}(s, g), captures the global relationship between the current state and the goal. We show that our bilinear decomposition scheme substantially improves data efficiency, and has superior transfer to out-of-distribution goals compared to prior methods. Empirical evidence is provided on the simulated Fetch robot task-suite and dexterous manipulation with a Shadow hand.

State-of-the-art deep Q-learning methods update Q-values using state transition tuples sampled from the experience replay buffer. This strategy often uniformly and randomly samples or prioritizes data sampling based on measures such as the temporal difference (TD) error. Such sampling strategies can be inefficient at learning Q-function because a state's Q-value depends on the Q-value of successor states. If the data sampling strategy ignores the precision of the Q-value estimate of the next state, it can lead to useless and often incorrect updates to the Q-values. To mitigate this issue, we organize the agent's experience into a graph that explicitly tracks the dependency between Q-values of states. Each edge in the graph represents a transition between two states by executing a single action. We perform value backups via a breadth-first search starting from that expands vertices in the graph starting from the set of terminal states and successively moving backward. We empirically show that our method is substantially more data-efficient than several baselines on a diverse range of goal-reaching tasks. Notably, the proposed method also outperforms baselines that consume more batches of training experience and operates from high-dimensional observational data such as images.

This paper introduces a novel approach to active feature acquisition for classification, which is the task of sequentially selecting the most informative subset of features to achieve optimal prediction performance during testing while minimizing cost. The proposed approach involves a new lazy model that is significantly faster and more efficient compared to existing methods, while still producing comparable accuracy results. During the test phase, the proposed approach utilizes Fisher scores for feature ranking to identify the most important feature at each step. In the next step the training dataset is filtered based on the observed value of the selected feature and then we continue this process to reach to acceptable accuracy or limit of the budget for feature acquisition. The performance of the proposed approach was evaluated on synthetic and real datasets, including our new synthetic dataset, CUBE dataset and also real dataset Forest. The experimental results demonstrate that our approach achieves competitive accuracy results compared to existing methods, while significantly outperforming them in terms of speed. The source code of the algorithm is released at github with this link: https://github.com/alimirzaei/FCwSFS.

In healthcare, there is much interest in estimating policies, or mappings from covariates to treatment decisions. Recently, there is also interest in constraining these estimated policies to the standard of care, which generated the observed data. A relative sparsity penalty was proposed to derive policies that have sparse, explainable differences from the standard of care, facilitating justification of the new policy. However, the developers of this penalty only considered estimation, not inference. Here, we develop inference for the relative sparsity objective function, because characterizing uncertainty is crucial to applications in medicine. Further, in the relative sparsity work, the authors only considered the single-stage decision case; here, we consider the more general, multi-stage case. Inference is difficult, because the relative sparsity objective depends on the unpenalized value function, which is unstable and has infinite estimands in the binary action case. Further, one must deal with a non-differentiable penalty. To tackle these issues, we nest a weighted Trust Region Policy Optimization function within a relative sparsity objective, implement an adaptive relative sparsity penalty, and propose a sample-splitting framework for post-selection inference. We study the asymptotic behavior of our proposed approaches, perform extensive simulations, and analyze a real, electronic health record dataset.

We present a framework for achieving end-to-end management objectives for multiple services that concurrently execute on a service mesh. We apply reinforcement learning (RL) techniques to train an agent that periodically performs control actions to reallocate resources. We develop and evaluate the framework using a laboratory testbed where we run information and computing services on a service mesh, supported by the Istio and Kubernetes platforms. We investigate different management objectives that include end-to-end delay bounds on service requests, throughput objectives, cost-related objectives, and service differentiation. We compute the control policies on a simulator rather than on the testbed, which speeds up the training time by orders of magnitude for the scenarios we study. Our proposed framework is novel in that it advocates a top-down approach whereby the management objectives are defined first and then mapped onto the available control actions. It allows us to execute several types of control actions simultaneously. By first learning the system model and the operating region from testbed traces, we can train the agent for different management objectives in parallel.

Trust-region methods based on Kullback-Leibler divergence are pervasively used to stabilize policy optimization in reinforcement learning. In this paper, we exploit more flexible metrics and examine two natural extensions of policy optimization with Wasserstein and Sinkhorn trust regions, namely Wasserstein policy optimization (WPO) and Sinkhorn policy optimization (SPO). Instead of restricting the policy to a parametric distribution class, we directly optimize the policy distribution and derive their closed-form policy updates based on the Lagrangian duality. Theoretically, we show that WPO guarantees a monotonic performance improvement, and SPO provably converges to WPO as the entropic regularizer diminishes. Moreover, we prove that with a decaying Lagrangian multiplier to the trust region constraint, both methods converge to global optimality. Experiments across tabular domains, robotic locomotion, and continuous control tasks further demonstrate the performance improvement of both approaches, more robustness of WPO to sample insufficiency, and faster convergence of SPO, over state-of-art policy gradient methods.

A novel Policy Gradient (PG) algorithm, called Matryoshka Policy Gradient (MPG), is introduced and studied, in the context of max-entropy reinforcement learning, where an agent aims at maximising entropy bonuses additional to its cumulative rewards. MPG differs from standard PG in that it trains a sequence of policies to learn finite horizon tasks simultaneously, instead of a single policy for the single standard objective. For softmax policies, we prove convergence of MPG and global optimality of the limit by showing that the only critical point of the MPG objective is the optimal policy; these results hold true even in the case of continuous compact state space. MPG is intuitive, theoretically sound and we furthermore show that the optimal policy of the standard max-entropy objective can be approximated arbitrarily well by the optimal policy of the MPG framework. Finally, we justify that MPG is well suited when the policies are parametrized with neural networks and we provide an simple criterion to verify the global optimality of the policy at convergence. As a proof of concept, we evaluate numerically MPG on standard test benchmarks.

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 \emph{all historical policies} does not necessarily benefit model prediction for the \emph{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 \textit{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.

Indirect simultaneous positioning (ISP), where internal tissue points are placed at desired locations indirectly through the manipulation of boundary points, is a type of subtask frequently performed in robotic surgeries. Although challenging due to complex tissue dynamics, automating the task can potentially reduce the workload of surgeons. This paper presents a sim-to-real framework for learning to automate the task without interacting with a real environment, and for planning preoperatively to find the grasping points that minimize local tissue deformation. A control policy is learned using deep reinforcement learning (DRL) in the FEM-based simulation environment and transferred to real-world situation. Grasping points are planned in the simulator by utilizing the trained policy using Bayesian optimization (BO). Inconsistent simulation performance is overcome by formulating the problem as a state augmented Markov decision process (MDP). Experimental results show that the learned policy places the internal tissue points accurately, and that the planned grasping points yield small tissue deformation among the trials. The proposed learning and planning scheme is able to automate internal tissue point manipulation in surgeries and has the potential to be generalized to complex surgical scenarios.

Offline Reinforcement Learning (RL), which seeks to perform standard RL tasks using a pre-existing dataset of interactions with an MDP, is a key frontier in the effort to make RL methods more widely applicable. The ability to incorporate existing data into RL algorithms is crucial in many promising application domains. In safety-critical areas, such as autonomous driving (Kiran et al., 2020), the randomized exploration that characterizes online algorithms is not ethically tolerable. Even in lower-stakes applications, such as advertising (Cai et al., 2017), naively adopting online algorithms could mean throwing away vast reserves of previouslycollected data. The development of efficient offline algorithms promises to broaden RL's applicability by allowing practitioners to exercise some much needed domain-specific control over the training process. Given a dataset, D, of interactions with an MDP M, two tasks that we may hope to achieve in offline RL are Offline Policy Evaluation (Yin & Wang, 2020) and Offline Learning (Lange et al., 2012). In Offline Policy Evaluation (OPE), we seek to estimate the value of a target policy π under M. In Offline Learning (OL), the goal is to use D to find a good policy π Π where Π is some policy class. In this paper, we largely focus on OPE.