BAM version 3 is in the final stages of testing and we managed to get an interview with the mastermind behind much of the modelling and validation that has gone into the new version; David Karlsson, Senior Mathematical Fatigue Risk Modelling Specialist at Jeppesen.
David, what can we expect from BAM 3, and when will we see this new release 'in action'?
- You should expect the same high performance as in BAM 2, a somewhat softer sleep prediction, improved modelling of task load, and improved sleep prediction in large time-zone transitions when physiology clashes with society. You will also find a novel approach for the modelling of acclimatization, which reduces the dependency on time zones and instead leans more heavily on light exposure. You can also expect the same leading validation methodology to be used as for BAM 2, but now using much more data.
Wow, that was quite a lot. Can you tell me more about the validation efforts? And again; when can we expect to start using this new release?
- Naturally, validation is of key importance in this work. The updated model must not only match collected actuals on an overall improved level, but also with good margin 'defend' the increased complexity when adding new mechanisms to the model. Jeppesen has today, I dare say, more collected fatigue data of high-quality than any other organisation, making these large-scale data-driven improvements quite viable - although not in any way simple. We need to be careful to neither over-inform, nor overfit the model to the collected data, and also ensure validation is done on data not being part of 'training' the model in the first place. A good methodology is already there, through our earlier work, which we are now piggy-backing on. When improving BAM we are in part using purpose-built parameter optimization software for identifying the best combination of settings for the internal model parameters in huge optimization jobs. What mainly remains now is to also double-check against predictions from BAM 2 on large sets of customer data and work sequences out of our pattern generator.
What do you mean by a pattern generator? And please, when will BAM 3 be released?
- Our pattern generator is a tool that produces an extreme amount, tens of millions, of airline-relevant crew scheduling patterns of generic nature such as combinations of early and late duties, but also patterns crossing time zones in a lot of different timing combinations with and without augmented crew etc. We are using this generator for 'regression tests' between model versions, with particular interest of course in the patterns that predicts quite differently between them. So far the results, also here, are very encouraging.
Thank you David. We will try to squeeze you for a few additional answers and some more details regarding the above in a coming issue of this newsletter. Hopefully also a prediction graph illustrating some differences between BAM 2 and 3. And possibly a date?
- OK. Let's see about that.
