The aim of multi-task reinforcement learning is two-fold: (1) efficiently learn by training against multiple tasks and (2) quickly adapt, using limited samples, to a variety of new tasks. In this work, the tasks correspond to reward functions for environments with the same (or similar) dynamical models. We propose to learn a dynamical model during the training process and use this model to perform sample-efficient adaptation to new tasks at test time. We use significantly fewer samples by performing policy optimization only in a "virtual" environment whose transitions are given by our learned dynamical model. Our algorithm sequentially trains against several tasks. Upon encountering a new task, we first warm-up a policy on our learned dynamical model, which requires no new samples from the environment. We then adapt the dynamical model with samples from this policy in the real environment. We evaluate our approach on several continuous control benchmarks and demonstrate its efficacy over MAML, a state-of-the-art meta-learning algorithm, on these tasks.
We study the problem of learning sequential decision-making policies in settings with multiple state-action representations. Such settings naturally arise in many domains, such as planning (e.g., multiple integer programming formulations) and various combinatorial optimization problems (e.g., those with both integer programming and graph-based formulations). Inspired by the classical co-training framework for classification, we study the problem of co-training for policy learning. We present sufficient conditions under which learning from two views can improve upon learning from a single view alone. Motivated by these theoretical insights, we present a meta-algorithm for co-training for sequential decision making. Our framework is compatible with both reinforcement learning and imitation learning. We validate the effectiveness of our approach across a wide range of tasks, including discrete/continuous control and combinatorial optimization.
Existing on-policy imitation learning algorithms, such as DAgger, assume access to a fixed supervisor. However, there are many settings where the supervisor may evolve during policy learning, such as a human performing a novel task or an improving algorithmic controller. We formalize imitation learning from a "converging supervisor" and provide sublinear static and dynamic regret guarantees against the best policy in hindsight with labels from the converged supervisor, even when labels during learning are only from intermediate supervisors. We then show that this framework is closely connected to a class of reinforcement learning (RL) algorithms known as dual policy iteration (DPI), which alternate between training a reactive learner with imitation learning and a model-based supervisor with data from the learner. Experiments suggest that when this framework is applied with the state-of-the-art deep model-based RL algorithm PETS as an improving supervisor, it outperforms deep RL baselines on continuous control tasks and provides up to an 80-fold speedup in policy evaluation.
Many deep reinforcement learning algorithms contain inductive biases that sculpt the agent's objective and its interface to the environment. These inductive biases can take many forms, including domain knowledge and pretuned hyper-parameters. In general, there is a trade-off between generality and performance when algorithms use such biases. Stronger biases can lead to faster learning, but weaker biases can potentially lead to more general algorithms. This trade-off is important because inductive biases are not free; substantial effort may be required to obtain relevant domain knowledge or to tune hyper-parameters effectively. In this paper, we re-examine several domain-specific components that bias the objective and the environmental interface of common deep reinforcement learning agents. We investigated whether the performance deteriorates when these components are replaced with adaptive solutions from the literature. In our experiments, performance sometimes decreased with the adaptive components, as one might expect when comparing to components crafted for the domain, but sometimes the adaptive components performed better. We investigated the main benefit of having fewer domain-specific components, by comparing the learning performance of the two systems on a different set of continuous control problems, without additional tuning of either system. As hypothesized, the system with adaptive components performed better on many of the new tasks.
Deep reinforcement learning (RL) algorithms can use high-capacity deep networks to learn directly from image observations. However, these high-dimensional observation spaces present a number of challenges in practice, since the policy must now solve two problems: representation learning and task learning. In this work, we tackle these two problems separately, by explicitly learning latent representations that can accelerate reinforcement learning from images. We propose the stochastic latent actor-critic (SLAC) algorithm: a sample-efficient and high-performing RL algorithm for learning policies for complex continuous control tasks directly from high-dimensional image inputs. SLAC provides a novel and principled approach for unifying stochastic sequential models and RL into a single method, by learning a compact latent representation and then performing RL in the model's learned latent space. Our experimental evaluation demonstrates that our method outperforms both model-free and model-based alternatives in terms of final performance and sample efficiency, on a range of difficult image-based control tasks. Our code and videos of our results are available at our website.
In recent years significant progress has been made in dealing with challenging problems using reinforcement learning.Despite its great success, reinforcement learning still faces challenge in continuous control tasks. Conventional methods always compute the derivatives of the optimal goal with a costly computation resources, and are inefficient, unstable and lack of robust-ness when dealing with such tasks. Alternatively, derivative-based methods treat the optimization process as a blackbox and show robustness and stability in learning continuous control tasks, but not data efficient in learning. The combination of both methods so as to get the best of the both has raised attention. However, most of the existing combination works adopt complex neural networks (NNs) as the policy for control. The double-edged sword of deep NNs can yield better performance, but also makes it difficult for parameter tuning and computation. To this end, in this paper we presents a novel method called FiDi-RL, which incorporates deep RL with Finite-Difference (FiDi) policy search.FiDi-RL combines Deep Deterministic Policy Gradients (DDPG)with Augment Random Search (ARS) and aims at improving the data efficiency of ARS. The empirical results show that FiDi-RL can improves the performance and stability of ARS, and provide competitive results against some existing deep reinforcement learning methods
Model-based reinforcement learning has the potential to be more sample efficient than model-free approaches. However, existing model-based methods are vulnerable to model bias, which leads to poor generalization and asymptotic performance compared to model-free counterparts. In addition, they are typically based on the model predictive control (MPC) framework, which not only is computationally inefficient at decision time but also does not enable policy transfer due to the lack of an explicit policy representation. In this paper, we propose a novel uncertainty-aware model-based policy optimization framework which solves those issues. In this framework, the agent simultaneously learns an uncertainty-aware dynamics model and optimizes the policy according to these learned models. In the optimization step, the policy gradient is computed by automatic differentiation through the models. With respect to sample efficiency alone, our approach shows promising results on challenging continuous control benchmarks with competitive asymptotic performance and significantly lower sample complexity than state-of-the-art baselines.
We consider the problem of imitation learning from expert demonstrations in partially observable Markov decision processes (POMDPs). Belief representations, which characterize the distribution over the latent states in a POMDP, have been modeled using recurrent neural networks and probabilistic latent variable models, and shown to be effective for reinforcement learning in POMDPs. In this work, we investigate the belief representation learning problem for generative adversarial imitation learning in POMDPs. Instead of training the belief module and the policy separately as suggested in prior work, we learn the belief module jointly with the policy, using a task-aware imitation loss to ensure that the representation is more aligned with the policy's objective. To improve robustness of representation, we introduce several informative belief regularization techniques, including multi-step prediction of dynamics and action-sequences. Evaluated on various partially observable continuous-control locomotion tasks, our belief-module imitation learning approach (BMIL) substantially outperforms several baselines, including the original GAIL algorithm and the task-agnostic belief learning algorithm. Extensive ablation analysis indicates the effectiveness of task-aware belief learning and belief regularization.
Deep reinforcement learning has made significant progress in the field of continuous control, such as physical control and autonomous driving. However, it is challenging for a reinforcement model to learn a policy for each task sequentially due to catastrophic forgetting. Specifically, the model would forget knowledge it learned in the past when trained on a new task. We consider this challenge from two perspectives: i) acquiring task-specific skills is difficult since task information and rewards are not highly related; ii) learning knowledge from previous experience is difficult in continuous control domains. In this paper, we introduce an end-to-end framework namely Continual Diversity Adversarial Network (CDAN). We first develop an unsupervised diversity exploration method to learn task-specific skills using an unsupervised objective. Then, we propose an adversarial self-correction mechanism to learn knowledge by exploiting past experience. The two learning procedures are presumably reciprocal. To evaluate the proposed method, we propose a new continuous reinforcement learning environment named Continual Ant Maze (CAM) and a new metric termed Normalized Shorten Distance (NSD). The experimental results confirm the effectiveness of diversity exploration and self-correction. It is worthwhile noting that our final result outperforms baseline by 18.35% in terms of NSD, and 0.61 according to the average reward.
Experience replay enables reinforcement learning agents to memorize and reuse past experiences, just as humans replay memories for the situation at hand. Contemporary off-policy algorithms either replay past experiences uniformly or utilize a rule-based replay strategy, which may be sub-optimal. In this work, we consider learning a replay policy to optimize the cumulative reward. Replay learning is challenging because the replay memory is noisy and large, and the cumulative reward is unstable. To address these issues, we propose a novel experience replay optimization (ERO) framework which alternately updates two policies: the agent policy, and the replay policy. The agent is updated to maximize the cumulative reward based on the replayed data, while the replay policy is updated to provide the agent with the most useful experiences. The conducted experiments on various continuous control tasks demonstrate the effectiveness of ERO, empirically showing promise in experience replay learning to improve the performance of off-policy reinforcement learning algorithms.
We consider deterministic Markov decision processes (MDPs) and apply max-plus algebra tools to approximate the value iteration algorithm by a smaller-dimensional iteration based on a representation on dictionaries of value functions. The setup naturally leads to novel theoretical results which are simply formulated due to the max-plus algebra structure. For example, when considering a fixed (non adaptive) finite basis, the computational complexity of approximating the optimal value function is not directly related to the number of states, but to notions of covering numbers of the state space. In order to break the curse of dimensionality in factored state-spaces, we consider adaptive basis that can adapt to particular problems leading to an algorithm similar to matching pursuit from signal processing. They currently come with no theoretical guarantees but work empirically well on simple deterministic MDPs derived from low-dimensional continuous control problems. We focus primarily on deterministic MDPs but note that the framework can be applied to all MDPs by considering measure-based formulations.
Extracting and predicting object structure and dynamics from videos without supervision is a major challenge in machine learning. To address this challenge, we adopt a keypoint-based image representation and learn a stochastic dynamics model of the keypoints. Future frames are reconstructed from the keypoints and a reference frame. By modeling dynamics in the keypoint coordinate space, we achieve stable learning and avoid compounding of errors in pixel space. Our method improves upon unstructured representations both for pixel-level video prediction and for downstream tasks requiring object-level understanding of motion dynamics. We evaluate our model on diverse datasets: a multi-agent sports dataset, the Human3.6M dataset, and datasets based on continuous control tasks from the DeepMind Control Suite. The spatially structured representation outperforms unstructured representations on a range of motion-related tasks such as object tracking, action recognition and reward prediction.
We provide a framework for incorporating robustness -- to perturbations in the transition dynamics which we refer to as model misspecification -- into continuous control Reinforcement Learning (RL) algorithms. We specifically focus on incorporating robustness into a state-of-the-art continuous control RL algorithm called Maximum a-posteriori Policy Optimization (MPO). We achieve this by learning a policy that optimizes for a worst case expected return objective and derive a corresponding robust entropy-regularized Bellman contraction operator. In addition, we introduce a less conservative, soft-robust, entropy-regularized objective with a corresponding Bellman operator. We show that both, robust and soft-robust policies, outperform their non-robust counterparts in nine Mujoco domains with environment perturbations. In addition, we show improved robust performance on a high-dimensional, simulated, dexterous robotic hand. Finally, we present multiple investigative experiments that provide a deeper insight into the robustness framework. This includes an adaptation to another continuous control RL algorithm as well as learning the uncertainty set from offline data. Performance videos can be found online at https://sites.google.com/view/robust-rl.
The outcome of Jacobian singular values regularization was studied for supervised learning problems. It also was shown that Jacobian conditioning regularization can help to avoid the ``mode-collapse'' problem in Generative Adversarial Networks. In this paper, we try to answer the following question: Can information about policy conditioning help to shape a more stable and general policy of reinforcement learning agents? To answer this question, we conduct a study of Jacobian conditioning behavior during policy optimization. To the best of our knowledge, this is the first work that research condition number in reinforcement learning agents. We propose a conditioning regularization algorithm and test its performance on the range of continuous control tasks. Finally, we compare algorithms on the CoinRun environment with separated train end test levels to analyze how conditioning regularization contributes to agents' generalization.
Exploration strategy design is one of the challenging problems in reinforcement learning~(RL), especially when the environment contains a large state space or sparse rewards. During exploration, the agent tries to discover novel areas or high reward~(quality) areas. In most existing methods, the novelty and quality in the neighboring area of the current state are not well utilized to guide the exploration of the agent. To tackle this problem, we propose a novel RL framework, called \underline{c}lustered \underline{r}einforcement \underline{l}earning~(CRL), for efficient exploration in RL. CRL adopts clustering to divide the collected states into several clusters, based on which a bonus reward reflecting both novelty and quality in the neighboring area~(cluster) of the current state is given to the agent. Experiments on a continuous control task and several \emph{Atari 2600} games show that CRL can outperform other state-of-the-art methods to achieve the best performance in most cases.
Lane change is a challenging task which requires delicate actions to ensure safety and comfort. Some recent studies have attempted to solve the lane-change control problem with Reinforcement Learning (RL), yet the action is confined to discrete action space. To overcome this limitation, we formulate the lane change behavior with continuous action in a model-free dynamic driving environment based on Deep Deterministic Policy Gradient (DDPG). The reward function, which is critical for learning the optimal policy, is defined by control values, position deviation status, and maneuvering time to provide the RL agent informative signals. The RL agent is trained from scratch without resorting to any prior knowledge of the environment and vehicle dynamics since they are not easy to obtain. Seven models under different hyperparameter settings are compared. A video showing the learning progress of the driving behavior is available. It demonstrates the RL vehicle agent initially runs out of road boundary frequently, but eventually has managed to smoothly and stably change to the target lane with a success rate of 100% under diverse driving situations in simulation.
Off-policy reinforcement learning aims to leverage experience collected from prior policies for sample-efficient learning. However, in practice, commonly used off-policy approximate dynamic programming methods based on Q-learning and actor-critic methods are highly sensitive to the data distribution, and can make only limited progress without collecting additional on-policy data. As a step towards more robust off-policy algorithms, we study the setting where the off-policy experience is fixed and there is no further interaction with the environment. We identify bootstrapping error as a key source of instability in current methods. Bootstrapping error is due to bootstrapping from actions that lie outside of the training data distribution, and it accumulates via the Bellman backup operator. We theoretically analyze bootstrapping error, and demonstrate how carefully constraining action selection in the backup can mitigate it. Based on our analysis, we propose a practical algorithm, bootstrapping error accumulation reduction (BEAR). We demonstrate that BEAR is able to learn robustly from different off-policy distributions, including random and suboptimal demonstrations, on a range of continuous control tasks.
We study the sample complexity of approximate policy iteration (PI) for the Linear Quadratic Regulator (LQR), building on a recent line of work using LQR as a testbed to understand the limits of reinforcement learning (RL) algorithms on continuous control tasks. Our analysis quantifies the tension between policy improvement and policy evaluation, and suggests that policy evaluation is the dominant factor in terms of sample complexity. Specifically, we show that to obtain a controller that is within $\varepsilon$ of the optimal LQR controller, each step of policy evaluation requires at most $(n+d)^3/\varepsilon^2$ samples, where $n$ is the dimension of the state vector and $d$ is the dimension of the input vector. On the other hand, only $\log(1/\varepsilon)$ policy improvement steps suffice, resulting in an overall sample complexity of $(n+d)^3 \varepsilon^{-2} \log(1/\varepsilon)$. We furthermore build on our analysis and construct a simple adaptive procedure based on $\varepsilon$-greedy exploration which relies on approximate PI as a sub-routine and obtains $T^{2/3}$ regret, improving upon a recent result of Abbasi-Yadkori et al.
Value functions are crucial for model-free Reinforcement Learning (RL) to obtain a policy implicitly or guide the policy updates. Value estimation heavily depends on the stochasticity of environmental dynamics and the quality of reward signals. In this paper, we propose a two-step understanding of value estimation from the perspective of future prediction, through decomposing the value function into a reward-independent future dynamics part and a policy-independent trajectory return part. We then derive a practical deep RL algorithm from the above decomposition, consisting of a convolutional trajectory representation model, a conditional variational dynamics model to predict the expected representation of future trajectory and a convex trajectory return model that maps a trajectory representation to its return. Our algorithm is evaluated in MuJoCo continuous control tasks and shows superior results under both common settings and delayed reward settings.
We observe that several existing policy gradient methods (such as vanilla policy gradient, PPO, A2C) may suffer from overly large gradients when the current policy is close to deterministic (even in some very simple environments), leading to an unstable training process. To address this issue, we propose a new method, called \emph{target distribution learning} (TDL), for policy improvement in reinforcement learning. TDL alternates between proposing a target distribution and training the policy network to approach the target distribution. TDL is more effective in constraining the KL divergence between updated policies, and hence leads to more stable policy improvements over iterations. Our experiments show that TDL algorithms perform comparably to (or better than) state-of-the-art algorithms for most continuous control tasks in the MuJoCo environment while being more stable in training.