We introduce the first temporal-difference learning algorithms that converge with smooth value function approximators, such as neural networks. Conventional temporal-difference (TD) methods, such as TD( \lambda), Q-learning and Sarsa have been used successfully with function approximation in many applications. However, it is well known that off-policy sampling, as well as nonlinear function approximation, can cause these algorithms to become unstable (i.e., the parameters of the approximator may diverge). Sutton et al (2009a,b) solved the problem of off-policy learning with linear TD algorithms by introducing a new objective function, related to the Bellman-error, and algorithms that perform stochastic gradient-descent on this function. In this paper, we generalize their work to nonlinear function approximation.

We provide new theoretical results for apprenticeship learning, a variant of reinforcement learning in which the true reward function is unknown, and the goal is to perform well relative to an observed expert. We study a common approach to learning from expert demonstrations: using a classification algorithm to learn to imitate the expert's behavior. Although this straightforward learning strategy is widely-used in practice, it has been subject to very little formal analysis. We prove that, if the learned classifier has error rate \eps, the difference between the value of the apprentice's policy and the expert's policy is O(\sqrt{\eps}) . Further, we prove that this difference is only O(\eps) when the expert's policy is close to optimal.

Likelihood ratio policy gradient methods have been some of the most successful reinforcement learning algorithms, especially for learning on physical systems. We describe how the likelihood ratio policy gradient can be derived from an importance sampling perspective. This derivation highlights how likelihood ratio methods under-use past experience by (a) using the past experience to estimate {\em only} the gradient of the expected return U(\theta) at the current policy parameterization \theta, rather than to obtain a more complete estimate of U(\theta), and (b) using past experience under the current policy {\em only} rather than using all past experience to improve the estimates. We present a new policy search method, which leverages both of these observations as well as generalized baselines---a new technique which generalizes commonly used baseline techniques for policy gradient methods. Our algorithm outperforms standard likelihood ratio policy gradient algorithms on several testbeds.

We study the problem of adaptive control of a high dimensional linear quadratic (LQ) system. Previous work established the asymptotic convergence to an optimal controller for various adaptive control schemes. More recently, an asymptotic regret bound of \tilde{O}(\sqrt{T}) was shown for T \gg p where p is the dimension of the state space. In this work we consider the case where the matrices describing the dynamic of the LQ system are sparse and their dimensions are large. We present an adaptive control scheme that for p \gg 1 and T \gg \polylog(p) achieves a regret bound of \tilde{O}(p \sqrt{T}) .

We present a novel l_1 regularized off-policy convergent TD-learning method (termed RO-TD), which is able to learn sparse representations of value functions with low computational complexity. The algorithmic framework underlying RO-TD integrates two key ideas: off-policy convergent gradient TD methods, such as TDC, and a convex-concave saddle-point formulation of non-smooth convex optimization, which enables first-order solvers and feature selection using online convex regularization. A detailed theoretical and experimental analysis of RO-TD is presented. A variety of experiments are presented to illustrate the off-policy convergence, sparse feature selection capability and low computational cost of the RO-TD algorithm.

We study the Constrained Convex Markov Decision Process (MDP), where the goal is to minimize a convex functional of the visitation measure, subject to a convex constraint. Designing algorithms for a constrained convex MDP faces several challenges, including (1) handling the large state space, (2) managing the exploration/exploitation tradeoff, and (3) solving the constrained optimization where the objective and the constraint are both nonlinear functions of the visitation measure. In this work, we present a model-based algorithm, Variational Primal-Dual Policy Optimization (VPDPO), in which Lagrangian and Fenchel duality are implemented to reformulate the original constrained problem into an unconstrained primal-dual optimization. Moreover, the primal variables are updated by model-based value iteration following the principle of Optimism in the Face of Uncertainty (OFU), while the dual variables are updated by gradient ascent. Moreover, by embedding the visitation measure into a finite-dimensional space, we can handle large state spaces by incorporating function approximation. Two notable examples are (1) Kernelized Nonlinear Regulators and (2) Low-rank MDPs. We prove that with an optimistic planning oracle, our algorithm achieves sublinear regret and constraint violation in both cases and can attain the globally optimal policy of the original constrained problem.

In this work, we address the problem of learning optimal behavior from sub-optimal datasets in the context of goal-conditioned offline reinforcement learning. To do so, we propose a novel way of approximating the optimal value function for goal-conditioned offline RL problems under sparse rewards, symmetric and deterministic actions. We study a property for representations to recover optimality and propose a new optimization objective that leads to such property. We use the learned value function to guide the learning of a policy in an actor-critic fashion, a method we name MetricRL. Experimentally, we show how our method consistently outperforms other offline RL baselines in learning from sub-optimal offline datasets. Moreover, we show the effectiveness of our method in dealing with high-dimensional observations and in multi-goal tasks.

Reinforcement Learning (RL) can effectively learn complex policies. However, learning these policies often demands extensive trial-and-error interactions with the environment. In many real-world scenarios, this approach is not practical due to the high costs of data collection and safety concerns. As a result, a common strategy is to transfer a policy trained in a low-cost, rapid source simulator to a real-world target environment. However, this process poses challenges. Simulators, no matter how advanced, cannot perfectly replicate the intricacies of the real world, leading to dynamics discrepancies between the source and target environments. Past research posited that the source domain must encompass all possible target transitions, a condition we term full support. However, expecting full support is often unrealistic, especially in scenarios where significant dynamics discrepancies arise. In this paper, our emphasis shifts to addressing large dynamics mismatch adaptation. We move away from the stringent full support condition of earlier research, focusing instead on crafting an effective policy for the target domain. Our proposed approach is simple but effective. It is anchored in the central concepts of the skewing and extension of source support towards target support to mitigate support deficiencies. Through comprehensive testing on a varied set of benchmarks, our method's efficacy stands out, showcasing notable improvements over previous techniques.

It is desirable in many multi-objective machine learning applications, such as multi-task learning with conflicting objectives and multi-objective reinforcement learning, to find a Pareto solution that can match a given preference of a decision maker. These problems are often large-scale with available gradient information but cannot be handled very well by the existing algorithms. To tackle this critical issue, this paper proposes a novel predict-and-correct framework for locating a Pareto solution that fits the preference of a decision maker. In the proposed framework, a constraint function is introduced in the search progress to align the solution with a user-specific preference, which can be optimized simultaneously with multiple objective functions. Experimental results show that our proposed method can efficiently find a particular Pareto solution under the demand of a decision maker for standard multiobjective benchmark, multi-task learning, and multi-objective reinforcement learning problems with more than thousands of decision variables. Code is available at: https://github.com/xzhang2523/pmgda. Our code is current provided in the pgmda.rar attached file and will be open-sourced after publication.}

Robotic manipulation of deformable linear objects (DLOs) is an active area of research, though emerging applications, like automotive wire harness installation, introduce constraints that have not been considered in prior work. Confined workspaces and limited visibility complicate prior assumptions of multi-robot manipulation and direct measurement of DLO configuration (state). This work focuses on single-arm manipulation of stiff DLOs (StDLOs) connected to form a DLO network (DLON), for which the measurements (output) are the endpoint poses of the DLON, which are subject to unknown dynamics during manipulation. To demonstrate feasibility of output-based control without state estimation, direct input-output dynamics are shown to exist by training neural network models on simulated trajectories. Output dynamics are then approximated with polynomials and found to contain well-known rigid body dynamics terms. A composite model consisting of a rigid body model and an online data-driven residual is developed, which predicts output dynamics more accurately than either model alone, and without prior experience with the system. An adaptive model predictive controller is developed with the composite model for DLON manipulation, which completes DLON installation tasks, both in simulation and with a physical automotive wire harness.