Many patients with respiratory illnesses need assistance to breathe normally. A mechanical ventilator uses pressure to help these patients. Usually, a nurse is needed to observe the patient’s condition and control the ventilator. Due to the global shortage of nurses, needing a nurse constantly to control a ventilator can be problematic during a pandemic. Therefore, building an AI agent to control the ventilator can be a remarkable help to healthcare systems. 

In this project, we build a reinforcement learning model based on "General Value Functions" to address this problem.

Agents in partially observable environments require state representations that incorporate historical information of their interactions with the environment. While recurrent neural networks (RNNs) are commonly used for this purpose, training them is computationally expensive. An alternative approach is to modify the agent's state construction, inspired by research on animal and human learning, which suggests that agents perform better when provided with a state representation containing information from past experiences. The current study extends this idea to fully observable environments and explores various methods of including a memory trace in the agent's state representation. The experiments conducted in a more complex environment demonstrate that some of these methods lead to improved performance, while others have limitations.