Adaptive average arterial pressure control by multi-agent on-policy reinforcement learning.

The current research introduces a model-free ultra-local model (MFULM) controller that utilizes the multi-agent on-policy reinforcement learning (MAOPRL) technique for remotely regulating blood pressure through precise drug dosing in a closed-loop system. Within the closed-loop system, there exists a MFULM controller, an observer, and an intelligent MAOPRL algorithm. Initially, a flexible MFULM controller is created to make adjustments to blood pressure and medication dosages. Following this, an observer is incorporated into the main controller to improve performance and stability by estimating states and disturbances. The controller parameters are optimized using MAOPRL in an adaptive manner, which involves the use of an actor-critic approach in an adaptive fashion. This approach enhances the adaptability of the controller by allowing for dynamic modifications to dosage and blood pressure control parameters. In the presence of disturbances or instabilities, the critic's feedback aids the actor in adjusting actions to reduce their impact, utilizing a complementary strategy to tackle deficiencies in the primary controller. Lastly, various evaluations, including assessments under normal conditions, adaptability between patients, and stability evaluations against mixed disturbances, have been carried out to confirm the efficiency and viability of the proposed method.