Reinforcement learning in the control of a simulated life support system
Quasny, Todd M
MetadataShow full item record
Since the 1970s, the National Aeronautics and Space Administration (NASA) has been conducting experiments to improve the duration and safety of manned space missions. For this purpose, an Advanced Life Support (ALS) system is being developed at NASA's Johnson Space Center (JSC). For research and testing purposes, an ALS system simulator, named BioSim, has been developed to simulate the interactions of the various subsystems of ALS. BioSim provides a testbed for researchers to develop and compare alternative control strategies for an ALS system. Reinforcement learning (RL) is a machine learning technique that finds effective control strategies. RL does this by interacting with the environment, and has been used successfully to control systems with noisy inputs and stochastic actions. RL methods are able to perform the real-time, reactive control that is vital in embedded control systems. This work demonstrates that reinforcement learning provides an excellent approach for finding an effective control policy for the water recovery subsystem of an ALS system. The control policy found by RL overcomes the inherent noisy inputs and stochastic actuation methods that exist in ALS systems. Using the policy found by RL, the mission duration is extended to nearly 450 days, at which point the mission ends for reasons other than the lack of consumable water. Since the mission does not end due to water concerns, it is concluded that an effective control policy for the water recovery system has been generated.