TOKYO, May 24, 2017 - Mitsubishi Electric Corporation (TOKYO: 6503) announced today that it has developed a proprietary deep-reinforcement algorithm for artificial-intelligence (AI) machine control that requires just one-fiftieth the number of trials compared to conventional AI control methods. The algorithm is expected to enable smart equipment such as industrial robots and vehicles to use sensors and cameras to rapidly learn about their environments for finely tuned AI-based control in unique environments.

Main Features

1)
Proprietary deep-reinforcement algorithm dramatically reduces learning time
- Machines achieve extra-smart deep-reinforcement learning using sensor and camera data
- Dramatically reduces number of trials and learning time compared to conventional deep-reinforcement learning methods
Conventional methods for AI-based smart work require extensive time to process huge amounts of data obtained from cameras and sensors, as well as extensive trials by machines using this data.
2)
Algorithm combined with Compact AI can be equipped in wide range of machines
- Combined with Mitsubishi Electric's Compact AI technology released in February 2016, the new algorithm requires just one-hundredth the amount of calculations compared to conventional methods
- Machines with limited processing resources can use the solution to perform deep-reinforcement learning
Working in combination with Mitsubishi Electric's Compact AI technology, the algorithm significantly reduces the calculation time compared to conventional methods, enabling deep-reinforcement learning to be deployed in a wide range of resource-limited equipment.
Solution Learning method Optimization time
New Fully automated machine learning Several minutes to 30 minutes
Existing Machine learning supported by human experts Several hours to half a day

Note that the releases are accurate at the time of publication but may be subject to change without notice.