Improving real time decisions Wednesday, 09 September 2015

American electrical engineers have patented a method that improves a controller’s ability to make real-time decisions.

Professor Frank Lewis from the University of Texas in Arlington, Draguna Vrabie from United Technologies and Kyriakos Vamvoudakis from the University of California in Santa Barbara have developed Integral Reinforcement Learning (IRL), a process by which a device learns and makes control decisions in reaction to a set of variables that changes based on each previous decision.

It involves learning by a batch process that requires taking a set of data prior to updating the control law. Their patent develops a new technology for IRL by which decisions can be made continuously in real time, online, allowing greater autonomy and faster response.

“Optimal feedback controllers allow a device to use the minimum energy necessary while saving time and fuel,” Lewis said.

“They are often seen in aircraft autopilots, emission controls in vehicles, and other similar technologies. The advantage of using adaptive integral reinforcement learning is that a device now can have optimal controls by looking within the system and calculating changes in real time, rather than offline where changes cannot be made until the device is no longer in use.”

Vamvoudakis added that adaptive integral reinforcement may eventually allow machines to work in increasingly difficult situations.

“Efficiency will be defined by the potential to adapt autonomously in decentralized, unknown and complex environments to enable capabilities beyond human limits,” he said.