Modular approaches and applications in reinforcement learning development and validation of next-generation frameworks

Abstract

Reinforcement learning (RL) provides a unique framework for addressing sequential decision-making problems. Despite the numerous software frameworks proposed to accelerate the development of new algorithms and applications, RL researchers and practitioners often still rely on custom code. This thesis identifies and addresses some core issues contributing to this trend. In the first part, we propose a modular approach for defining distributed RL schemes using basic, reusable building blocks. In the second part, we contribute to the creation of TorchRL, the official PyTorch domain library for general decision-making. TorchRL is designed to be efficient, scalable, and broadly applicable. Finally, we leverage and validate TorchRL by developing ACEGEN, a library for language-based generative drug discovery, and use it to explore new solutions in this field.