Improving Reliability and Adaptability via Program Steering

Software systems often contain several discrete modes of operation and a mechanism for switching between modes. Even when a multi-mode system has been tested by its developer, it may be unreliable in the field, because the developer cannot foresee and test for every possible scenario; unexpected situations in which the program fails or underperforms (for example, by choosing a non-optimal mode) are certain to arise. This research mitigates such problems by creating adaptive modal programs that handle unanticipated scenarios by autonomously selecting an appropriate mode. The technique creates a new mode selector via machine learning by training on good behavior in anticipated situations. The new controller can augment or replace the old one. Preliminary experiments indicate that our technique can re-derive ideal controllers and can improve the reliability and performance of good ones.