• No results found

2. Related Work

2.1. Occupant Location Prediction

2.1.2. Context-Based Prediction

The above approaches both consider predicting occupants’ future locations, but have a limited ability to use extra context elements when making a prediction. Context is important as it defines the ability of an approach to make different predictions in different circumstances. Future location prediction approaches which can’t take any extra context elements will always predict that an occupant will be in the same place at the same time of day. In fact the occupant’s identity and the time of day are themselves context elements – as demonstrated by the zeroth order Markov model discussed above, without the time of day element only the most frequent location for the occupant is predicted. Without knowing the occupant, only the most frequent location in the dataset will be predicted. Just as it is necessary to know the occupant’s identity and the target time of day in order to make a good prediction, as the correct prediction depends on these factors, it is also important to be able to include other elements of context which influence the occupants’ locations. Various extra elements of context have been considered in

15

existing approaches which focus on incorporating extra context elements into their predictions.

One such approach is (Petzold et al. 2005) a Bayesian Network approach is used to predict where an occupant will go next, when, and how long they will stay there for. The approach models a time slice (of arbitrary duration), which represents a stay in a single location, as a Bayesian network. The nodes in the network are the current room, current duration (i.e. length of stay), time of day and day of week, with current duration being dependent on the other three nodes. An occupant’s sequence of movements then becomes a sequence of length of identical networks, where is the number of previous locations which are considered. In each time slice the current room is dependent on the time of day and day of week of the previous time slice, and on the current room in all previous time slices. In this way the model predicts an occupant’s next location based on their recent locations, but also the current time of day and the day of the week. The duration of stay in both the current and predicted time slice are predicted solely off the location in the same time slice.

Figure 2-2 – Bayesian network from (Petzold et al. 2005)

As the structure of the network is fixed, the prediction of the occupant’s location is always dependent on the time of day and the day of the week, as well as their current location, which means that these elements of context must always be used. To change the context used, the structure of the network would have to be

16

restructured and the model retrained. Similarly the number of time slices used is fixed, resulting in all predictions being made using a fixed number of previous locations. As the model can train while predicting, location prediction accuracy with and without training on a training set was considered. For a sequence length of two, making predictions only when the occupant is not in their office, the approach achieved overall accuracy of 65% without training and 72% with training. Stay duration prediction was tested with networks including different combinations of parent nodes, and with and without training data. Using current room and time of day, making predictions only when the occupant is not in their own office, the approach achieved 70% accuracy without training and 78% with training.

In (Koehler et al. 2014) an ensemble approach is presented which combines the predictions of four classifiers (decision tree, 3-nearest neighbour, support vector machine and gradient boost) to independently predict when an occupant will transition to a new location and what location that will be. The approach uses a variety of elements of context: current location, time of arrival at current location, minutes passed since arriving at current location, current time, day of week, arrival time in the building, number of significant locations visited thus far today, previous two significant locations, duration of stay at previous significant location and time taken to transition to current location. The approach uses ten minute timeslots and significant locations are those which the occupant frequently remains in for at least one timeslot.

In order to determine which features from this list should be used when making a prediction the approach uses sequential floating forward selection (SFFS), which repeatedly add features to the feature set in order to improve accuracy, removing those which can be removed without negatively affecting accuracy. SFFS is applied independently for each classifier in order to find and use the best feature set for each. This process is repeated after each predicted day to react to changes in occupants’ routines.

As mentioned, the approach predicts where an occupant will transition to and when. Specifically, for a given number of minutes, it predicts whether the occupant

17

will stay in their current location for that duration, and if they will not, then it predicts where they will go. When asked whether the occupant will transition, each classifier responds with 1(yes) or 0 (no). If the mean prediction is equal or greater than 0.5 the overall prediction is 1, otherwise it is 0. If it is 1, then the classifiers each vote on the location, and the majority vote is the overall prediction, with random selection used as the tiebreaker. As the temporal and spatial predictions are made separately, accuracy is calculated for a two step prediction. To be correct, the algorithm must first correctly predict whether the occupant will stay where they are. If the occupant leaves, it must also correctly predict where they go. The approach was tested for look-aheads of 10 to 90 minutes, achieving over 90% accuracy for 10 minute look-ahead, dropping monotonically to just below 80% for 90 minutes.