Robust and Efficient Model-based Reinforcement Learning

强化学习（RL）关注的是训练数据驱动的代理来做出决策。特别是，与环境交互的强化学习代理需要学习最优策略，即在不同状态下采取哪些行动以最大化其回报。最近，强化学习已经成为机器学习中最突出的领域之一，因为强化学习方法在解决各个领域（例如，自动驾驶、核聚变、医疗保健、硬件设计等）的复杂任务方面具有巨大的潜力。然而，许多挑战仍然阻碍着它的广泛采用。当代强化学习算法通常是数据密集型的，缺乏鲁棒性保证。已建立的（深度）强化学习方法需要在某些环境（如电子游戏）中随时可用的大量数据。在数据采集成本很高的现实世界任务中，情况往往不是这样。另一个主要挑战是在确保可靠、健壮和安全的性能的同时，在现实世界中使用学习到的控制策略。本研究旨在提供具有严格统计和鲁棒性保证的实用的基于模型的强化学习算法。这在获取数据昂贵的安全关键应用中非常重要，例如，在核聚变中，通过昂贵的模拟器来学习控制等离子体的策略。关键的新颖之处在于将多用途的鲁棒性方面整合到基于模型的强化学习中，从而允许其在不同的应用程序和领域中广泛应用。这个项目的重点是设计算法，利用强大的非线性统计模型来了解世界，并可以处理现代强化学习任务中存在的大状态空间。重点是获得接近最优的策略，这些策略可以抵御环境动态的分布变化、（对抗性）数据损坏/异常值，并满足勘探过程中与应用程序相关的安全约束。一个主要的贡献将是对设计算法的新颖严格的统计样本复杂性保证，这些算法具有收敛到最优鲁棒和安全策略的特征。所获得的保证在独立于状态数的意义上是有效的，因此适用于复杂的应用。这需要为流行的统计模型设计新的稳健估计器和置信区间。此外，该项目将产生一个具有分布转移和攻击策略的整个测试平台，将提供基准测试标准和新型鲁棒强化学习算法的鲁棒性。该项目将是实现MBRL的鲁棒性和效率的第一个贡献，它提供了实用的算法，可以很容易地应用于新兴的有影响力的现实世界任务，如核等离子体的鲁棒控制（通向可持续能源的令人兴奋和有前途的道路）和芯片系统设计的有效发现。

Reinforcement learning (RL) is concerned with training data-driven agents to make decisions. In particular, an RL agent interacting with an environment needs to learn an optimal policy, i.e., which actions to take in different states to maximize its rewards. Recently, RL has become one of the most prominent areas of machine learning since RL methods can have tremendous potential in solving complex tasks across various fields (e.g., in autonomous driving, nuclear fusion, healthcare, hardware design, etc.). However, a number of challenges still stand in the way of its widespread adoption. Contemporary RL algorithms are often data-intensive and lack robustness guarantees. Established (deep) RL approaches require a vast amount of data that is readily available in some environments (e.g., in video games). This is often not the case with real-world tasks where data acquisition is costly. Another major challenge is to use the learned control policies in the real world while ensuring reliable, robust, and safe performance. This research aims to provide practical model-based RL algorithms with rigorous statistical and robustness guarantees. This is significant in safety-critical applications where obtaining data is expensive, e.g., in nuclear fusion, learning policies to control plasmas is performed via expensive simulators. The key novelty will be to incorporate the versatile robustness aspects into model-based RL allowing for its broad application across different applications and domains.This project focuses on designing algorithms that make use of powerful non-linear statistical models to learn about the world and can tackle large state spaces present in modern RL tasks. The focus is on obtaining near-optimal policies that are robust against distributional shifts in the environmental dynamics, (adversarial) data corruptions/outliers, and satisfy application-dependent safety constraints during exploration. A major contribution will be novel rigorous statistical sample complexity guarantees for designed algorithms that characterize convergence to optimal robust and safe policies. The obtained guarantees will be efficient in the sense of being independent of the number of states, and hence applicable to complex applications. This will require designing new robust estimators and confidence intervals for popular statistical models. Moreover, the project will result in an entire testbed with distributional shifts and attacking strategies that will be provided to benchmark the robustness of standard and novel robust RL algorithms. This project will be among the first contribution to achieving both robustness and efficiency in MBRL by providing practical algorithms that can be readily applied to emerging impactful real-world tasks such as robust control of nuclear plasmas (an exciting and promising path toward sustainable energy) and efficient discovery of system-on-chip designs.