|Application domain:||Lift systems|
|Further Specification:||Data sets|
|Pointers:||Contact Luc De Raedt (Luc.DeRaedt@cs.kuleuven.ac.be)|
|Data complexity estimation:||1000 examples|
This application is about learning rules for an expert system that tries to optimise the performance of a lift system (a set of elevators), e.g. minimise waiting times, minimise energy consumption, ...
One dataset was generated artificially, two others were obtained from real lift systems.
From the data rules are induced that try to predict whether there will be much traffic in the near future, and between which floors this traffic will occur. For instance, in most office buildings there is much upward traffic in the morning, and downward traffic in the evening; if there is suddenly an increase in traffic to one specific floor, there may be a meeting there and it can be expected that more people will want to go to that floor in the near future. This shows that the lift system should be able to react dynamically to a changing environment. Therefore rules that relate traffic to the time of day (for instance) do not suffice. Instead, rules are induced that try to predict traffic for a small time period in the near future, based on the traffic observed during a small time interval in the recent past.
The performance of a lift control systems that makes use of the induced rules has been compared with that of several existing lift control systems (programmed by domain experts). To this aim a simulator was used that simulates the behaviour of the control system in the artificial and real-world environments from which the data were obtained. Depending on the environment, using the learned rules made it possible to reduce the waiting times by 7% to 30%, with respect to the best manually-coded system.