Model-free Reinforcement Learning for Model-based Control: Towards Safe, Interpretable and Sample-efficient Agents

Training sophisticated agents for optimal decision-making under uncertainty has been key to the rapid development of modern autonomous systems across fields. Notably, model-free reinforcement learning (RL) has enabled decision-making agents to improve their performance directly through system interactions, with minimal prior knowledge about the system. Yet, model-free RL has generally relied on agents equipped with deep neural network function approximators, appealing to the networks' expressivity to capture the agent's policy and value function for complex systems. However, neural networks amplify the issues of sample inefficiency, unsafe learning, and limited interpretability in model-free RL. To this end, this work introduces model-based agents as a compelling alternative for control policy approximation, leveraging adaptable models of system dynamics, cost, and constraints for safe policy learning. These models can encode prior system knowledge to inform, constrain, and aid in explaining the agent's decisions, while deficiencies due to model mismatch can be remedied with model-free RL. We outline the benefits and challenges of learning model-based agents -- exemplified by model predictive control -- and detail the primary learning approaches: Bayesian optimization, policy search RL, and offline strategies, along with their respective strengths. While model-free RL has long been established, its interplay with model-based agents remains largely unexplored, motivating our perspective on their combined potentials for sample-efficient learning of safe and interpretable decision-making agents.

MPCritic: A plug-and-play MPC architecture for reinforcement learning

The reinforcement learning (RL) and model predictive control (MPC) communities have developed vast ecosystems of theoretical approaches and computational tools for solving optimal control problems. Given their conceptual similarities but differing strengths, there has been increasing interest in synergizing RL and MPC. However, existing approaches tend to be limited for various reasons, including computational cost of MPC in an RL algorithm and software hurdles towards seamless integration of MPC and RL tools. These challenges often result in the use of "simple" MPC schemes or RL algorithms, neglecting the state-of-the-art in both areas. This paper presents MPCritic, a machine learning-friendly architecture that interfaces seamlessly with MPC tools. MPCritic utilizes the loss landscape defined by a parameterized MPC problem, focusing on "soft" optimization over batched training steps; thereby updating the MPC parameters while avoiding costly minimization and parametric sensitivities. Since the MPC structure is preserved during training, an MPC agent can be readily used for online deployment, where robust constraint satisfaction is paramount. We demonstrate the versatility of MPCritic, in terms of MPC architectures and RL algorithms that it can accommodate, on classic control benchmarks.

Accelerating Black-Box Molecular Property Optimization by Adaptively Learning Sparse Subspaces

Molecular property optimization (MPO) problems are inherently challenging since they are formulated over discrete, unstructured spaces and the labeling process involves expensive simulations or experiments, which fundamentally limits the amount of available data. Bayesian optimization (BO) is a powerful and popular framework for efficient optimization of noisy, black-box objective functions (e.g., measured property values), thus is a potentially attractive framework for MPO. To apply BO to MPO problems, one must select a structured molecular representation that enables construction of a probabilistic surrogate model. Many molecular representations have been developed, however, they are all high-dimensional, which introduces important challenges in the BO process -- mainly because the curse of dimensionality makes it difficult to define and perform inference over a suitable class of surrogate models. This challenge has been recently addressed by learning a lower-dimensional encoding of a SMILE or graph representation of a molecule in an unsupervised manner and then performing BO in the encoded space. In this work, we show that such methods have a tendency to "get stuck," which we hypothesize occurs since the mapping from the encoded space to property values is not necessarily well-modeled by a Gaussian process. We argue for an alternative approach that combines numerical molecular descriptors with a sparse axis-aligned Gaussian process model, which is capable of rapidly identifying sparse subspaces that are most relevant to modeling the unknown property function. We demonstrate that our proposed method substantially outperforms existing MPO methods on a variety of benchmark and real-world problems. Specifically, we show that our method can routinely find near-optimal molecules out of a set of more than $>100$k alternatives within 100 or fewer expensive queries.