Daniel Rangel Martinez

Machine learning techniques have become an attractive alternative to solve optimization problems using historical data and simulations. Supervised and unsupervised learning, which can produce models for forecasting or analysis, have been useful tools for the development of mathematical models. On the cases where data is not available or when the system is subject to continuous changes, there is the alternative to use Deep Reinforcement Learning (DRL) methods. These methods aim to create intelligent agents that can take decisions and learn how to perform tasks, e.g., optimal schedules and control profiles. My research focuses on the implementation of DRL algorithms as optimization methods in chemical engineering scenarios, specifically scheduling problems under uncertainty on job-shop machine environments. The capacity of DRL agents to adapt decisions to changes and to be pre-trained with historical data (using supervised learning methods) makes them an attractive tool to handle this type of dynamic processes. On and off-policy algorithms like Deep Deterministic Policy Gradient (DDPG) and Proximal Policy Optimization (PPO) are state-of-art models that I use to perform these implementations. The objective is to create efficient DRL agents that can handle the effects of uncertainty in on-line scheduling processes.