Inferring ECA-based rules for ambient intelligence using evolutionary feature extraction

One of the goals in Ambient Intelligence is to enable Intelligent Environments to take decisions based on the perceived context. In our previous work, we successfully explored how the inhabitants can communicate their own preferences with the environment using Event-Condition-Action (ECA) rules. The easiness of the communication language combined with an appropriate explanation mechanism gives trust to the Intelligent Environment actions. However, defining every preference, and maintaining them up-to-date can be cumbersome. Therefore, a complementary mechanism is required to learn from user behavior and adapt to small changes without being explicitly requested for. Inferring behaviors effectively from data collected from sensors in an Intelligent Environment is a challenging problem. The main issues include primitive representation of data, the necessity of a high number of sensors, and dealing with few training data collected in a short time. We present MFE3/GA^DR, an evolutionary constructive induction method to ease inferring inhabitants' preferences from data collected from simple sensors. We show that this method detects successfully relevant sensors and constructs highly informative features that abstract relations among them. The constructed features, in addition to improving significantly the learning accuracy, break down and encapsulate the performance of inhabitants into decision trees that can easily be converted to ECA rules for further use in the Intelligent Environment. Comparing the empirical results show that our method can reduce a large set of complex ECA rules that represent the preferences to a smaller set of simple ECA rules.