Executive Summary : | Process systems are always under tremendous pressure to operate by maximizing profit and productivity while preserving safety and reliability. However, this is a challenging task due to (i)nonlinear dynamics, (ii) unsteady-state operation, and (iii) demanding quality requirements of the products. Model predictive control has been traditionally being used to achieve these objectives. An accurate process model and online model maintenance are two major impediments that always need to be overcome. The advent of artificial intelligence and machine learning opened up new vistas that could be adopted in the context of process control for making them "intelligent". Therefore, there is a great incentive if the control strategy can interact directly with the process trajectory and provide a control solution for online course modification through learning operational data profiles. Reinforcement learning (RL) can be a potential solution in this context. In contrast to traditional controllers, RL-based controllers do not require a process model or control rules. Instead, learn about process dynamics by interacting directly with the operating environment. As a result, controller performance is no longer dependent on high fidelity process models. Moreover, unlike traditional controllers, RL-based controllers learn from experience and past and improve control policies at every step. In this proposal, we examine and propose extensions for a specific class of RL architectures known as actor-critic methods for making them amenable for process control. Based on the understanding gained by PI from our previous projects, herein, we propose the following objectives. (i)Using inverse RL framework to compute reward functions that can capture performance metrics, (ii) Improving actor-critic performance by using both deterministic and stochastic actors, (iii) Development of weighted actor based multi-actor RL, (iv) Actor-critic frameworks of objective (2) and/or (3) with sharing experience, (v)Simulation-based validations of process systems |