Internet of Things (IoT) applications showcase all the quirkiness of the real world: Sensors break, actuators get swapped and the world changes in unpredictable ways. A traditional learning approach of train-test-release for deploying models doesn't fit well. In addition, learning the relationship between actuator commands and resulting sensor changes is beyond the capabilities of standard naive classification and regression algorithms. One solution to this is to use an adaptive, model-based reinforcement learning algorithm, such as the one in BECCA 7. This overcomes many of the barriers to making the world around us smart.

YouTube video
O'Reilly interview about adaptive reinforcement learning for IoT
PDF slides [4MB]

I presented this during the Strata+Hadoop World Conference at San Jose.