Learning policies from fixed offline datasets is a key challenge to scale up reinforcement learning (RL) algorithms towards practical applications. This is often because off-policy RL algorithms suffer from distributional shift, due to mismatch between dataset and the target policy, leading to high variance and over-estimation of value functions. In this work, we propose variance regularization for offline RL algorithms, using stationary distribution corrections. We show that by using Fenchel duality, we can avoid double sampling issues for computing the gradient of the variance regularizer. The proposed algorithm for offline variance regularization (OVAR) can be used to augment any existing offline policy optimization algorithms. We show that the regularizer leads to a lower bound to the offline policy optimization objective, which can help avoid over-estimation errors, and explains the benefits of our approach across a range of continuous control domains when compared to existing state-of-the-art algorithms.
In recent years distributional reinforcement learning has produced many state of the art results. Increasingly sample efficient Distributional algorithms for the discrete action domain have been developed over time that vary primarily in the way they parameterize their approximations of value distributions, and how they quantify the differences between those distributions. In this work we transfer three of the most well-known and successful of those algorithms (QR-DQN, IQN and FQF) to the continuous action domain by extending two powerful actor-critic algorithms (TD3 and SAC) with distributional critics. We investigate whether the relative performance of the methods for the discrete action space translates to the continuous case. To that end we compare them empirically on the pybullet implementations of a set of continuous control tasks. Our results indicate qualitative invariance regarding the number and placement of distributional atoms in the deterministic, continuous action setting.
Deep Reinforcement Learning is emerging as a promising approach for the continuous control task of robotic arm movement. However, the challenges of learning robust and versatile control capabilities are still far from being resolved for real-world applications, mainly because of two common issues of this learning paradigm: the exploration strategy and the slow learning speed, sometimes known as "the curse of dimensionality". This work aims at exploring and assessing the advantages of the application of Quantum Computing to one of the state-of-art Reinforcement Learning techniques for continuous control - namely Soft Actor-Critic. Specifically, the performance of a Variational Quantum Soft Actor-Critic on the movement of a virtual robotic arm has been investigated by means of digital simulations of quantum circuits. A quantum advantage over the classical algorithm has been found in terms of a significant decrease in the amount of required parameters for satisfactory model training, paving the way for further promising developments.
A default assumption in reinforcement learning (RL) and optimal control is that observations arrive at discrete time points on a fixed clock cycle. Yet, many applications involve continuous-time systems where the time discretization, in principle, can be managed. The impact of time discretization on RL methods has not been fully characterized in existing theory, but a more detailed analysis of its effect could reveal opportunities for improving data-efficiency. We address this gap by analyzing Monte-Carlo policy evaluation for LQR systems and uncover a fundamental trade-off between approximation and statistical error in value estimation. Importantly, these two errors behave differently to time discretization, leading to an optimal choice of temporal resolution for a given data budget. These findings show that managing the temporal resolution can provably improve policy evaluation efficiency in LQR systems with finite data. Empirically, we demonstrate the trade-off in numerical simulations of LQR instances and standard RL benchmarks for non-linear continuous control.
Cross-Entropy Method (CEM) is commonly used for planning in model-based reinforcement learning (MBRL) where a centralized approach is typically utilized to update the sampling distribution based on only the top-$k$ operation's results on samples. In this paper, we show that such a centralized approach makes CEM vulnerable to local optima, thus impairing its sample efficiency. To tackle this issue, we propose Decentralized CEM (DecentCEM), a simple but effective improvement over classical CEM, by using an ensemble of CEM instances running independently from one another, and each performing a local improvement of its own sampling distribution. We provide both theoretical and empirical analysis to demonstrate the effectiveness of this simple decentralized approach. We empirically show that, compared to the classical centralized approach using either a single or even a mixture of Gaussian distributions, our DecentCEM finds the global optimum much more consistently thus improves the sample efficiency. Furthermore, we plug in our DecentCEM in the planning problem of MBRL, and evaluate our approach in several continuous control environments, with comparison to the state-of-art CEM based MBRL approaches (PETS and POPLIN). Results show sample efficiency improvement by simply replacing the classical CEM module with our DecentCEM module, while only sacrificing a reasonable amount of computational cost. Lastly, we conduct ablation studies for more in-depth analysis. Code is available at https://github.com/vincentzhang/decentCEM
The policy gradient method enjoys the simplicity of the objective where the agent optimizes the cumulative reward directly. Moreover, in the continuous action domain, parameterized distribution of action distribution allows easy control of exploration, resulting from the variance of the representing distribution. Entropy can play an essential role in policy optimization by selecting the stochastic policy, which eventually helps better explore the environment in reinforcement learning (RL). However, the stochasticity often reduces as the training progresses; thus, the policy becomes less exploratory. Additionally, certain parametric distributions might only work for some environments and require extensive hyperparameter tuning. This paper aims to mitigate these issues. In particular, we propose an algorithm called Robust Policy Optimization (RPO), which leverages a perturbed distribution. We hypothesize that our method encourages high-entropy actions and provides a way to represent the action space better. We further provide empirical evidence to verify our hypothesis. We evaluated our methods on various continuous control tasks from DeepMind Control, OpenAI Gym, Pybullet, and IsaacGym. We observed that in many settings, RPO increases the policy entropy early in training and then maintains a certain level of entropy throughout the training period. Eventually, our agent RPO shows consistently improved performance compared to PPO and other techniques: entropy regularization, different distributions, and data augmentation. Furthermore, in several settings, our method stays robust in performance, while other baseline mechanisms fail to improve and even worsen the performance.
Proximal Policy Optimization (PPO) is a highly popular policy-based deep reinforcement learning (DRL) approach. However, we observe that the homogeneous exploration process in PPO could cause an unexpected stability issue in the training phase. To address this issue, we propose PPO-UE, a PPO variant equipped with self-adaptive uncertainty-aware explorations (UEs) based on a ratio uncertainty level. The proposed PPO-UE is designed to improve convergence speed and performance with an optimized ratio uncertainty level. Through extensive sensitivity analysis by varying the ratio uncertainty level, our proposed PPO-UE considerably outperforms the baseline PPO in Roboschool continuous control tasks.
Inferring reward functions from human behavior is at the center of value alignment - aligning AI objectives with what we, humans, actually want. But doing so relies on models of how humans behave given their objectives. After decades of research in cognitive science, neuroscience, and behavioral economics, obtaining accurate human models remains an open research topic. This begs the question: how accurate do these models need to be in order for the reward inference to be accurate? On the one hand, if small errors in the model can lead to catastrophic error in inference, the entire framework of reward learning seems ill-fated, as we will never have perfect models of human behavior. On the other hand, if as our models improve, we can have a guarantee that reward accuracy also improves, this would show the benefit of more work on the modeling side. We study this question both theoretically and empirically. We do show that it is unfortunately possible to construct small adversarial biases in behavior that lead to arbitrarily large errors in the inferred reward. However, and arguably more importantly, we are also able to identify reasonable assumptions under which the reward inference error can be bounded linearly in the error in the human model. Finally, we verify our theoretical insights in discrete and continuous control tasks with simulated and human data.
In classic reinforcement learning algorithms, agents make decisions at discrete and fixed time intervals. The duration between decisions becomes a crucial hyperparameter, as setting it too short may increase the problem's difficulty by requiring the agent to make numerous decisions to achieve its goal while setting it too long can result in the agent losing control over the system. However, physical systems do not necessarily require a constant control frequency, and for learning agents, it is often preferable to operate with a low frequency when possible and a high frequency when necessary. We propose a framework called Continuous-Time Continuous-Options (CTCO), where the agent chooses options as sub-policies of variable durations. These options are time-continuous and can interact with the system at any desired frequency providing a smooth change of actions. We demonstrate the effectiveness of CTCO by comparing its performance to classical RL and temporal-abstraction RL methods on simulated continuous control tasks with various action-cycle times. We show that our algorithm's performance is not affected by the choice of environment interaction frequency. Furthermore, we demonstrate the efficacy of CTCO in facilitating exploration in a real-world visual reaching task for a 7 DOF robotic arm with sparse rewards.
Deep reinforcement learning (DRL) provides a new way to generate robot control policy. However, the process of training control policy requires lengthy exploration, resulting in a low sample efficiency of reinforcement learning (RL) in real-world tasks. Both imitation learning (IL) and learning from demonstrations (LfD) improve the training process by using expert demonstrations, but imperfect expert demonstrations can mislead policy improvement. Offline to Online reinforcement learning requires a lot of offline data to initialize the policy, and distribution shift can easily lead to performance degradation during online fine-tuning. To solve the above problems, we propose a learning from demonstrations method named A-SILfD, which treats expert demonstrations as the agent's successful experiences and uses experiences to constrain policy improvement. Furthermore, we prevent performance degradation due to large estimation errors in the Q-function by the ensemble Q-functions. Our experiments show that A-SILfD can significantly improve sample efficiency using a small number of different quality expert demonstrations. In four Mujoco continuous control tasks, A-SILfD can significantly outperform baseline methods after 150,000 steps of online training and is not misled by imperfect expert demonstrations during training.
Go-Explore achieved breakthrough performance on challenging reinforcement learning (RL) tasks with sparse rewards. The key insight of Go-Explore was that successful exploration requires an agent to first return to an interesting state ('Go'), and only then explore into unknown terrain ('Explore'). We refer to such exploration after a goal is reached as 'post-exploration'. In this paper, we present a clear ablation study of post-exploration in a general intrinsically motivated goal exploration process (IMGEP) framework, that the Go-Explore paper did not show. We study the isolated potential of post-exploration, by turning it on and off within the same algorithm under both tabular and deep RL settings on both discrete navigation and continuous control tasks. Experiments on a range of MiniGrid and Mujoco environments show that post-exploration indeed helps IMGEP agents reach more diverse states and boosts their performance. In short, our work suggests that RL researchers should consider to use post-exploration in IMGEP when possible since it is effective, method-agnostic and easy to implement.
Many real-world continuous control problems are in the dilemma of weighing the pros and cons, multi-objective reinforcement learning (MORL) serves as a generic framework of learning control policies for different preferences over objectives. However, the existing MORL methods either rely on multiple passes of explicit search for finding the Pareto front and therefore are not sample-efficient, or utilizes a shared policy network for coarse knowledge sharing among policies. To boost the sample efficiency of MORL, we propose Q-Pensieve, a policy improvement scheme that stores a collection of Q-snapshots to jointly determine the policy update direction and thereby enables data sharing at the policy level. We show that Q-Pensieve can be naturally integrated with soft policy iteration with convergence guarantee. To substantiate this concept, we propose the technique of Q replay buffer, which stores the learned Q-networks from the past iterations, and arrive at a practical actor-critic implementation. Through extensive experiments and an ablation study, we demonstrate that with much fewer samples, the proposed algorithm can outperform the benchmark MORL methods on a variety of MORL benchmark tasks.
This paper proposes a novel model-based policy gradient algorithm for tracking dynamic targets using a mobile robot, equipped with an onboard sensor with limited field of view. The task is to obtain a continuous control policy for the mobile robot to collect sensor measurements that reduce uncertainty in the target states, measured by the target distribution entropy. We design a neural network control policy with the robot $SE(3)$ pose and the mean vector and information matrix of the joint target distribution as inputs and attention layers to handle variable numbers of targets. We also derive the gradient of the target entropy with respect to the network parameters explicitly, allowing efficient model-based policy gradient optimization.
Deep Reinforcement Learning (DRL) has the potential to be used for synthesizing feedback controllers (agents) for various complex systems with unknown dynamics. These systems are expected to satisfy diverse safety and liveness properties best captured using temporal logic. In RL, the reward function plays a crucial role in specifying the desired behaviour of these agents. However, the problem of designing the reward function for an RL agent to satisfy complex temporal logic specifications has received limited attention in the literature. To address this, we provide a systematic way of generating rewards in real-time by using the quantitative semantics of Signal Temporal Logic (STL), a widely used temporal logic to specify the behaviour of cyber-physical systems. We propose a new quantitative semantics for STL having several desirable properties, making it suitable for reward generation. We evaluate our STL-based reinforcement learning mechanism on several complex continuous control benchmarks and compare our STL semantics with those available in the literature in terms of their efficacy in synthesizing the controller agent. Experimental results establish our new semantics to be the most suitable for synthesizing feedback controllers for complex continuous dynamical systems through reinforcement learning.
This paper studies the control problem for safety-critical multi-agent systems based on quadratic programming (QP). Each controlled agent is modeled as a cascade connection of an integrator and an uncertain nonlinear actuation system. In particular, the integrator represents the position-velocity relation, and the actuation system describes the dynamic response of the actual velocity to the velocity reference signal. The notion of input-to-output stability (IOS) is employed to characterize the essential velocity-tracking capability of the actuation system. The uncertain actuation dynamics may cause infeasibility or discontinuous solutions of QP algorithms for collision avoidance. Also, the interaction between the controlled integrator and the uncertain actuation dynamics may lead to significant robustness issues. By using nonlinear control methods and numerical optimization methods, this paper first contributes a new feasible-set reshaping technique and a refined QP algorithm for feasibility, robustness, and local Lipschitz continuity. Then, we present a nonlinear small-gain analysis to handle the inherent interaction for guaranteed safety of the closed-loop multi-agent system. The proposed methods are illustrated by numerical simulations and a physical experiment.
Neural algorithmic reasoning studies the problem of learning algorithms with neural networks, especially with graph architectures. A recent proposal, XLVIN, reaps the benefits of using a graph neural network that simulates the value iteration algorithm in deep reinforcement learning agents. It allows model-free planning without access to privileged information about the environment, which is usually unavailable. However, XLVIN only supports discrete action spaces, and is hence nontrivially applicable to most tasks of real-world interest. We expand XLVIN to continuous action spaces by discretization, and evaluate several selective expansion policies to deal with the large planning graphs. Our proposal, CNAP, demonstrates how neural algorithmic reasoning can make a measurable impact in higher-dimensional continuous control settings, such as MuJoCo, bringing gains in low-data settings and outperforming model-free baselines.
Intrinsic motivation is a promising exploration technique for solving reinforcement learning tasks with sparse or absent extrinsic rewards. There exist two technical challenges in implementing intrinsic motivation: 1) how to design a proper intrinsic objective to facilitate efficient exploration; and 2) how to combine the intrinsic objective with the extrinsic objective to help find better solutions. In the current literature, the intrinsic objectives are all designed in a task-agnostic manner and combined with the extrinsic objective via simple addition (or used by itself for reward-free pre-training). In this work, we show that these designs would fail in typical sparse-reward continuous control tasks. To address the problem, we propose Constrained Intrinsic Motivation (CIM) to leverage readily attainable task priors to construct a constrained intrinsic objective, and at the same time, exploit the Lagrangian method to adaptively balance the intrinsic and extrinsic objectives via a simultaneous-maximization framework. We empirically show, on multiple sparse-reward continuous control tasks, that our CIM approach achieves greatly improved performance and sample efficiency over state-of-the-art methods. Moreover, the key techniques of our CIM can also be plugged into existing methods to boost their performances.
Decision Transformer (DT) is a recently proposed architecture for Reinforcement Learning that frames the decision-making process as an auto-regressive sequence modeling problem and uses a Transformer model to predict the next action in a sequence of states, actions, and rewards. In this paper, we analyze how crucial the Transformer model is in the complete DT architecture on continuous control tasks. Namely, we replace the Transformer by an LSTM model while keeping the other parts unchanged to obtain what we call a Decision LSTM model. We compare it to DT on continuous control tasks, including pendulum swing-up and stabilization, in simulation and on physical hardware. Our experiments show that DT struggles with continuous control problems, such as inverted pendulum and Furuta pendulum stabilization. On the other hand, the proposed Decision LSTM is able to achieve expert-level performance on these tasks, in addition to learning a swing-up controller on the real system. These results suggest that the strength of the Decision Transformer for continuous control tasks may lie in the overall sequential modeling architecture and not in the Transformer per se.
In this paper, hypernetworks are trained to generate behaviors across a range of unseen task conditions, via a novel TD-based training objective and data from a set of near-optimal RL solutions for training tasks. This work relates to meta RL, contextual RL, and transfer learning, with a particular focus on zero-shot performance at test time, enabled by knowledge of the task parameters (also known as context). Our technical approach is based upon viewing each RL algorithm as a mapping from the MDP specifics to the near-optimal value function and policy and seek to approximate it with a hypernetwork that can generate near-optimal value functions and policies, given the parameters of the MDP. We show that, under certain conditions, this mapping can be considered as a supervised learning problem. We empirically evaluate the effectiveness of our method for zero-shot transfer to new reward and transition dynamics on a series of continuous control tasks from DeepMind Control Suite. Our method demonstrates significant improvements over baselines from multitask and meta RL approaches.
Offline multi-agent reinforcement learning (MARL) aims to learn effective multi-agent policies from pre-collected datasets, which is an important step toward the deployment of multi-agent systems in real-world applications. However, in practice, each individual behavior policy that generates multi-agent joint trajectories usually has a different level of how well it performs. e.g., an agent is a random policy while other agents are medium policies. In the cooperative game with global reward, one agent learned by existing offline MARL often inherits this random policy, jeopardizing the performance of the entire team. In this paper, we investigate offline MARL with explicit consideration on the diversity of agent-wise trajectories and propose a novel framework called Shared Individual Trajectories (SIT) to address this problem. Specifically, an attention-based reward decomposition network assigns the credit to each agent through a differentiable key-value memory mechanism in an offline manner. These decomposed credits are then used to reconstruct the joint offline datasets into prioritized experience replay with individual trajectories, thereafter agents can share their good trajectories and conservatively train their policies with a graph attention network (GAT) based critic. We evaluate our method in both discrete control (i.e., StarCraft II and multi-agent particle environment) and continuous control (i.e, multi-agent mujoco). The results indicate that our method achieves significantly better results in complex and mixed offline multi-agent datasets, especially when the difference of data quality between individual trajectories is large.