SHMUC prototype > D_Behaviour_Pattern_Abnormalities
One of the major challenges of identifying behaviour patterns is that their appearance in terms of sensor patterns can vary immensely between people, and even within the activities of one person. For example, there are many different ways to cook dinner, depending upon what it is; evidence from various smart home datasets suggests that between 4 and 58 actions are needed depending upon the type of food prepared, and other behaviours exhibit similar variation. This makes one area where it can be particularly difficult to decide what a smart home should be able to detect, and how to balance the risk of false positives. We use two scenarios for the common and easily understood task of making a cup of tea to illustrate the complexity of identifying errors even in simple task.
D1 Making Tea With Sugar
D2 Making Tea With Cold Water
D3 Making Tea With Hot And Cold Water
D4 Taking A Shower While Cooking
These use cases demonstrate one of the most difficult aspects of behaviour recognition, the massive potential variation in behaviour presentation and the difference between a ‘safe’ one and one that demonstrates illness can be subtle. These use cases also highlight some differences between using logic-based methods and probabilistic-based methods. In the first scenario, the order in which the tea making activities occur differed from that usually seen. This is perfectly reasonable, but it can be a challenge for the smart home to recognise it depending on how it represents behaviours. A method based on a Markov model may well not even notice that there is a difference, while a lookup table that expects things to occur in a particular order would not be able to learn this without significant additional storage. However, the description in the scenario suggests that the system should be able to identify that this is a trivial change to a learnt behaviour and therefore modify its representation itself. This is difficult to do without including some kind of reasoning system that can use some knowledge base to identify whether or not the order matters; clearly removing the cup from the cupboard after the water has been poured would not be valid. Rather than patterns, which can become rather cumbersome when different orderings need to be allowed, it could be that the system can use its knowledge base to identify that the essential ingredients of a cup of tea include tea and hot water for dissolving tea, and that sugar is not essential, but it can be added at any stage because it does not affect the other actions.
So adding sugar could be considered as a ‘normal’ action in this activity. The system can then automatically update the tea-making activity pattern to include this. The use of knowledge and reasoning is important in this situation. An alternative could be to use partial orders for the patterns.
This problem can be simplified by focusing upon the final state of the sequence rather than tracing the whole progress. In such case, the use of AI planning would be worth. However, the smart home needs to detect the abnormal behaviour when it is occurring, not when the system reaches the final state. So, the use of AI planning requires some related problems solved, that will be discussed later in the discussion section.