| Abstract | A multi-mode software system contains several distinct modes of operation and a controller for deciding when to switch between modes. Even when developers rigorously test a multi-mode system before deployment, they cannot foresee and test for every possible usage scenario. As a result, unexpected situations in which the program fails or underperforms (for example, by choosing a non-optimal mode) may arise. This research aims to mitigate such problems by creating a new mode selector that examines the current situation, then chooses a mode that has been successful in the past, in situations like the current one. The technique, called program steering, creates a new mode selector via machine learning from good behavior in testing or in successful operation. Such a strategy, which generalizes the knowledge that a programmer has built into the system, may select an appropriate mode even when the original controller cannot. We have performed experiments on robot control programs written in a month-long programming competition. Augmenting these programs via our program steering technique had a substantial positive effect on their performance in new environments. |
