Model-based reinforcement learning (RL) enjoys several benefits, such as data-efficiency and planning, by learning a model of the environment's dynamics. However, learning a global model that can generalize across different dynamics is a challenging task. To tackle this problem, we decompose the task of learning a global dynamics model into two stages: (a) learning a context latent vector that captures the local dynamics, then (b) predicting the next state conditioned on it. In order to encode dynamics-specific information into the context latent vector, we introduce a novel loss function that encourages the context latent vector to be useful for predicting both forward and backward dynamics. The proposed method achieves superior generalization ability across various simulated robotics and control tasks, compared to existing RL schemes.
Action delays degrade the performance of reinforcement learning in many real-world systems. This paper proposes a formal definition of delay-aware Markov Decision Process and proves it can be transformed into standard MDP with augmented states using the Markov reward process. We develop a delay-aware model-based reinforcement learning framework that can incorporate the multi-step delay into the learned system models without learning effort. Experiments with the Gym and MuJoCo platforms show that the proposed delay-aware model-based algorithm is more efficient in training and transferable between systems with various durations of delay compared with off-policy model-free reinforcement learning methods. Codes available at: https://github.com/baimingc/dambrl.
Our brain receives a dynamically changing stream of sensorimotor data. Yet, we perceive a rather organized world, which we segment into and perceive as events. Computational theories of cognitive science on event-predictive cognition suggest that our brain forms generative, event-predictive models by segmenting sensorimotor data into suitable chunks of contextual experiences. Here, we introduce a hierarchical, surprise-gated recurrent neural network architecture, which models this process and develops compact compressions of distinct event-like contexts. The architecture contains a contextual LSTM layer, which develops generative compressions of ongoing and subsequent contexts. These compressions are passed into a GRU-like layer, which uses surprise signals to update its recurrent latent state. The latent state is passed forward into another LSTM layer, which processes actual dynamic sensory flow in the light of the provided latent, contextual compression signals. Our model shows to develop distinct event compressions and achieves the best performance on multiple event processing tasks. The architecture may be very useful for the further development of resource-efficient learning, hierarchical model-based reinforcement learning, as well as the development of artificial event-predictive cognition and intelligence.
Human-computer interactive systems that rely on machine learning are becoming paramount to the lives of millions of people who use digital assistants on a daily basis. Yet, further advances are limited by the availability of data and the cost of acquiring new samples. One way to address this problem is by improving the sample efficiency of current approaches. As a solution path, we present a model-based reinforcement learning algorithm for an interactive dialogue task. We build on commonly used actor-critic methods, adding an environment model and planner that augments a learning agent to learn the model of the environment dynamics. Our results show that, on a simulation that mimics the interactive task, our algorithm requires 70 times fewer samples, compared to the baseline of commonly used model-free algorithm, and demonstrates 2~times better performance asymptotically. Moreover, we introduce a novel contribution of computing a soft planner policy and further updating a model-free policy yielding a less computationally expensive model-free agent as good as the model-based one. This model-based architecture serves as a foundation that can be extended to other human-computer interactive tasks allowing further advances in this direction.
Identifying algorithms that flexibly and efficiently discover temporally-extended multi-phase plans is an essential step for the advancement of robotics and model-based reinforcement learning. The core problem of long-range planning is finding an efficient way to search through the tree of possible action sequences. Existing non-learned planning solutions from the Task and Motion Planning (TAMP) literature rely on the existence of logical descriptions for the effects and preconditions for actions. This constraint allows TAMP methods to efficiently reduce the tree search problem but limits their ability to generalize to unseen and complex physical environments. In contrast, deep reinforcement learning (DRL) methods use flexible neural-network-based function approximators to discover policies that generalize naturally to unseen circumstances. However, DRL methods struggle to handle the very sparse reward landscapes inherent to long-range multi-step planning situations. Here, we propose the Curious Sample Planner (CSP), which fuses elements of TAMP and DRL by combining a curiosity-guided sampling strategy with imitation learning to accelerate planning. We show that CSP can efficiently discover interesting and complex temporally-extended plans for solving a wide range of physically realistic 3D tasks. In contrast, standard planning and learning methods often fail to solve these tasks at all or do so only with a huge and highly variable number of training samples. We explore the use of a variety of curiosity metrics with CSP and analyze the types of solutions that CSP discovers. Finally, we show that CSP supports task transfer so that the exploration policies learned during experience with one task can help improve efficiency on related tasks.
Although recent model-free reinforcement learning algorithms have been shown to be capable of mastering complicated decision-making tasks, the sample complexity of these methods has remained a hurdle to utilizing them in many real-world applications. In this regard, model-based reinforcement learning proposes some remedies. Yet, inherently, model-based methods are more computationally expensive and susceptible to sub-optimality. One reason is that model-generated data are always less accurate than real data, and this often leads to inaccurate transition and reward function models. With the aim to mitigate this problem, this work presents the notion of survival by discussing cases in which the agent's goal is to survive and its analogy to maximizing the expected rewards. To that end, a substitute model for the reward function approximator is introduced that learns to avoid terminal states rather than to maximize accumulated rewards from safe states. Focusing on terminal states, as a small fraction of state-space, reduces the training effort drastically. Next, a model-based reinforcement learning method is proposed (Survive) to train an agent to avoid dangerous states through a safety map model built upon temporal credit assignment in the vicinity of terminal states. Finally, the performance of the presented algorithm is investigated, along with a comparison between the proposed and current methods.
Recent works in high-dimensional model-predictive control and model-based reinforcement learning with learned dynamics and reward models have resorted to population-based optimization methods, such as the Cross-Entropy Method (CEM), for planning a sequence of actions. To decide on an action to take, CEM conducts a search for the action sequence with the highest return according to the dynamics model and reward. Action sequences are typically randomly sampled from an unconditional Gaussian distribution and evaluated on the environment. This distribution is iteratively updated towards action sequences with higher returns. However, this planning method can be very inefficient, especially for high-dimensional action spaces. An alternative line of approaches optimize action sequences directly via gradient descent, but are prone to local optima. We propose a method to solve this planning problem by interleaving CEM and gradient descent steps in optimizing the action sequence. Our experiments show faster convergence of the proposed hybrid approach, even for high-dimensional action spaces, avoidance of local minima, and better or equal performance to CEM. Code accompanying the paper is available here https://github.com/homangab/gradcem.
Model-based reinforcement learning (MBRL) has recently gained immense interest due to its potential for sample efficiency and ability to incorporate off-policy data. However, designing stable and efficient MBRL algorithms using rich function approximators have remained challenging. To help expose the practical challenges in MBRL and simplify algorithm design from the lens of abstraction, we develop a new framework that casts MBRL as a game between: (1) a policy player, which attempts to maximize rewards under the learned model; (2) a model player, which attempts to fit the real-world data collected by the policy player. For algorithm development, we construct a Stackelberg game between the two players, and show that it can be solved with approximate bi-level optimization. This gives rise to two natural families of algorithms for MBRL based on which player is chosen as the leader in the Stackelberg game. Together, they encapsulate, unify, and generalize many previous MBRL algorithms. Furthermore, our framework is consistent with and provides a clear basis for heuristics known to be important in practice from prior works. Finally, through experiments we validate that our proposed algorithms are highly sample efficient, match the asymptotic performance of model-free policy gradient, and scale gracefully to high-dimensional tasks like dexterous hand manipulation. Additional details and code can be obtained from the project page at https://sites.google.com/view/mbrl-game
Applying reinforcement learning to robotic systems poses a number of challenging problems. A key requirement is the ability to handle continuous state and action spaces while remaining within a limited time and resource budget. Additionally, for safe operation, the system must make robust decisions under hard constraints. To address these challenges, we propose a model based approach that combines Gaussian Process regression and Receding Horizon Control. Using sparse spectrum Gaussian Processes, we extend previous work by updating the dynamics model incrementally from a stream of sensory data. This results in an agent that can learn and plan in real-time under non-linear constraints. We test our approach on a cart pole swing-up environment and demonstrate the benefits of online learning on an autonomous racing task. The environment's dynamics are learned from limited training data and can be reused in new task instances without retraining.
Access to a fast and easily copied forward model of a game is essential for model-based reinforcement learning and for algorithms such as Monte Carlo tree search, and is also beneficial as a source of unlimited experience data for model-free algorithms. Learning forward models is an interesting and important challenge in order to address problems where a model is not available. Building upon previous work on the Neural GPU, this paper introduces the Neural Game Engine, as a way to learn models directly from pixels. The learned models are able to generalise to different size game levels to the ones they were trained on without loss of accuracy. Results on 10 deterministic General Video Game AI games demonstrate competitive performance, with many of the games models being learned perfectly both in terms of pixel predictions and reward predictions. The pre-trained models are available through the OpenAI Gym interface and are available publicly for future research here: \url{https://github.com/Bam4d/Neural-Game-Engine}
Learning to control robots without requiring engineered models has been a long-term goal, promising diverse and novel applications. Yet, reinforcement learning has only achieved limited impact on real-time robot control due to its high demand of real-world interactions. In this work, by leveraging a learnt probabilistic model of drone dynamics, we learn a thrust-attitude controller for a quadrotor through model-based reinforcement learning. No prior knowledge of the flight dynamics is assumed; instead, a sequential latent variable model, used generatively and as an online filter, is learnt from raw sensory input. The controller and value function are optimised entirely by propagating stochastic analytic gradients through generated latent trajectories. We show that "learning to fly" can be achieved with less than 30 minutes of experience with a single drone, and can be deployed solely using onboard computational resources and sensors, on a self-built drone.
Collaboration requires agents to align their goals on the fly. Underlying the human ability to align goals with other agents is their ability to predict the intentions of others and actively update their own plans. We propose hierarchical predictive planning (HPP), a model-based reinforcement learning method for decentralized multiagent rendezvous. Starting with pretrained, single-agent point to point navigation policies and using noisy, high-dimensional sensor inputs like lidar, we first learn via self-supervision motion predictions of all agents on the team. Next, HPP uses the prediction models to propose and evaluate navigation subgoals for completing the rendezvous task without explicit communication among agents. We evaluate HPP in a suite of unseen environments, with increasing complexity and numbers of obstacles. We show that HPP outperforms alternative reinforcement learning, path planning, and heuristic-based baselines on challenging, unseen environments. Experiments in the real world demonstrate successful transfer of the prediction models from sim to real world without any additional fine-tuning. Altogether, HPP removes the need for a centralized operator in multiagent systems by combining model-based RL and inference methods, enabling agents to dynamically align plans.
Meta-learning algorithms can accelerate the model-based reinforcement learning (MBRL) algorithms by finding an initial set of parameters for the dynamical model such that the model can be trained to match the actual dynamics of the system with only a few data-points. However, in the real world, a robot might encounter any situation starting from motor failures to finding itself in a rocky terrain where the dynamics of the robot can be significantly different from one another. In this paper, first, we show that when meta-training situations (the prior situations) have such diverse dynamics, using a single set of meta-trained parameters as a starting point still requires a large number of observations from the real system to learn a useful model of the dynamics. Second, we propose an algorithm called FAMLE that mitigates this limitation by meta-training several initial starting points (i.e., initial parameters) for training the model and allows the robot to select the most suitable starting point to adapt the model to the current situation with only a few gradient steps. We compare FAMLE to MBRL, MBRL with a meta-trained model with MAML, and model-free policy search algorithm PPO for various simulated and real robotic tasks, and show that FAMLE allows the robots to adapt to novel damages in significantly fewer time-steps than the baselines.
This paper considers the problem of learning a model in model-based reinforcement learning (MBRL). We examine how the planning module of an MBRL algorithm uses the model, and propose that the model learning module should incorporate the way the planner is going to use the model. This is in contrast to conventional model learning approaches, such as those based on maximum likelihood estimate, that learn a predictive model of the environment without explicitly considering the interaction of the model and the planner. We focus on policy gradient type of planning algorithms and derive new loss functions for model learning that incorporate how the planner uses the model. We call this approach Policy-Aware Model Learning (PAML). We theoretically analyze a generic model-based policy gradient algorithm and provide a convergence guarantee for the optimized policy. We also empirically evaluate PAML on some benchmark problems, showing promising results.
In the present paper, we propose an extension of the Deep Planning Network (PlaNet), also referred to as PlaNet of the Bayesians (PlaNet-Bayes). There has been a growing demand in model predictive control (MPC) in partially observable environments in which complete information is unavailable because of, for example, lack of expensive sensors. PlaNet is a promising solution to realize such latent MPC, as it is used to train state-space models via model-based reinforcement learning (MBRL) and to conduct planning in the latent space. However, recent state-of-the-art strategies mentioned in MBRR literature, such as involving uncertainty into training and planning, have not been considered, significantly suppressing the training performance. The proposed extension is to make PlaNet uncertainty-aware on the basis of Bayesian inference, in which both model and action uncertainty are incorporated. Uncertainty in latent models is represented using a neural network ensemble to approximately infer model posteriors. The ensemble of optimal action candidates is also employed to capture multimodal uncertainty in the optimality. The concept of the action ensemble relies on a general variational inference MPC (VI-MPC) framework and its instance, probabilistic action ensemble with trajectory sampling (PaETS). In this paper, we extend VI-MPC and PaETS, which have been originally introduced in previous literature, to address partially observable cases. We experimentally compare the performances on continuous control tasks, and conclude that our method can consistently improve the asymptotic performance compared with PlaNet.
In this paper, we propose a derivative-free model learning framework for Reinforcement Learning (RL) algorithms based on Gaussian Process Regression (GPR). In many mechanical systems, only positions can be measured by the sensing instruments. Then, instead of representing the system state as suggested by the physics with a collection of positions, velocities, and accelerations, we define the state as the set of past position measurements. However, the equation of motions derived by physical first principles cannot be directly applied in this framework, being functions of velocities and accelerations. For this reason, we introduce a novel derivative-free physically-inspired kernel, which can be easily combined with nonparametric derivative-free Gaussian Process models. Tests performed on two real platforms show that the considered state definition combined with the proposed model improves estimation performance and data-efficiency w.r.t. traditional models based on GPR. Finally, we validate the proposed framework by solving two RL control problems for two real robotic systems.
Model-based reinforcement learning has been empirically demonstrated as a successful strategy to improve sample efficiency. In particular, Dyna is an elegant model-based architecture integrating learning and planning that provides huge flexibility of using a model. One of the most important components in Dyna is called search-control, which refers to the process of generating state or state-action pairs from which we query the model to acquire simulated experiences. Search-control is critical in improving learning efficiency. In this work, we propose a simple and novel search-control strategy by searching high frequency regions of the value function. Our main intuition is built on Shannon sampling theorem from signal processing, which indicates that a high frequency signal requires more samples to reconstruct. We empirically show that a high frequency function is more difficult to approximate. This suggests a search-control strategy: we should use states from high frequency regions of the value function to query the model to acquire more samples. We develop a simple strategy to locally measure the frequency of a function by gradient and hessian norms, and provide theoretical justification for this approach. We then apply our strategy to search-control in Dyna, and conduct experiments to show its property and effectiveness on benchmark domains.
Model-based reinforcement learning (MBRL) has been shown to be a powerful framework for data-efficiently learning control of continuous tasks. Recent work in MBRL has mostly focused on using more advanced function approximators and planning schemes, with little development of the general framework. In this paper, we identify a fundamental issue of the standard MBRL framework -- what we call the objective mismatch issue. Objective mismatch arises when one objective is optimized in the hope that a second, often uncorrelated, metric will also be optimized. In the context of MBRL, we characterize the objective mismatch between training the forward dynamics model w.r.t.~the likelihood of the one-step ahead prediction, and the overall goal of improving performance on a downstream control task. For example, this issue can emerge with the realization that dynamics models effective for a specific task do not necessarily need to be globally accurate, and vice versa globally accurate models might not be sufficiently accurate locally to obtain good control performance on a specific task. In our experiments, we study this objective mismatch issue and demonstrate that the likelihood of one-step ahead predictions is not always correlated with control performance. This observation highlights a critical limitation in the MBRL framework which will require further research to be fully understood and addressed. We propose an initial method to mitigate the mismatch issue by re-weighting dynamics model training. Building on it, we conclude with a discussion about other potential directions of research for addressing this issue.
Model-Based Reinforcement Learning (MBRL) offers a promising direction for sample efficient learning, often achieving state of the art results for continuous control tasks. However, many existing MBRL methods rely on combining greedy policies with exploration heuristics, and even those which utilize principled exploration bonuses construct dual objectives in an ad hoc fashion. In this paper we introduce Ready Policy One (RP1), a framework that views MBRL as an active learning problem, where we aim to improve the world model in the fewest samples possible. RP1 achieves this by utilizing a hybrid objective function, which crucially adapts during optimization, allowing the algorithm to trade off reward v.s. exploration at different stages of learning. In addition, we introduce a principled mechanism to terminate sample collection once we have a rich enough trajectory batch to improve the model. We rigorously evaluate our method on a variety of continuous control tasks, and demonstrate statistically significant gains over existing approaches.
Learning-based control aims to construct models of a system to use for planning or trajectory optimization, e.g. in model-based reinforcement learning. In order to obtain guarantees of safety in this context, uncertainty must be accurately quantified. This uncertainty may come from errors in learning (due to a lack of data, for example), or may be inherent to the system. Propagating uncertainty forward in learned dynamics models is a difficult problem. In this work we use deep learning to obtain expressive and flexible models of how distributions of trajectories behave, which we then use for nonlinear Model Predictive Control (MPC). We introduce a deep quantile regression framework for control that enforces probabilistic quantile bounds and quantifies epistemic uncertainty. Using our method we explore three different approaches for learning tubes that contain the possible trajectories of the system, and demonstrate how to use each of them in a Tube MPC scheme. We prove these schemes are recursively feasible and satisfy constraints with a desired margin of probability. We present experiments in simulation on a nonlinear quadrotor system, demonstrating the practical efficacy of these ideas.