I’m participating in a discussion later this week organised by CSIRO’s Data61. I’ll be talking about AI for decision support and optimisation. This is very different to large language models. AI solutions for decision support are formulated rather than trained. This reflects the practical realities of complex decision making.
- We may not want to merely replicate current practice. Although AI may replace a human, he or she may not be doing the best that could be done. There are opportunties to automate AND improve performance.
- Complex decision in areas like supply chain often rely on data that is controlled or provided by others who don’t have the same incentives to get it right
- Hopefully there is a lot less data on the outcomes of bad decisions than those of good decisions, so the training data is almost always incomplete or biased
We had a good example of potential problems at one of our customers. Some data analysts were looking at the operation and couldn’t make sense of the data. We looked at the data and the problem was clear. Drivers were doing deliveries all day and then confirming them at the end of the day. On the face of it, they were doing all their work in 10 minutes at the end of their shift!
For reasons like this, we formulate a problem that AI algorithms can solve, rather than just throwing data at the algorithm and hoping for the best. Why discard domain knowledge and engineering principles?
