We are frequently asked if optimization is AI or not. From an engineer’s perspective it’s a very uninteresting question – if you’re sitting in a thing with four wheels that runs on petrol and transports people to work, then it doesn’t really matter what you call it, you’ll end up at the office anyway.
On the other hand, the categories, where we put things, do really matter. They’re part of our mental landscape and guide our preconceptions about what we can expect from a thing belonging to a certain reference group. If you put a thing in the wrong category, people end up confused.
So in what category does optimization belong?
There are a lot of think-pieces out there about AI and you’ve probably read a number of them yourself. Typical AI features (from the data analytics point of view) are:
- It helps automate something previously manual
- It is self-learning
- It requires a lot of computation resources
- It gets better the more (high-quality) data you have
- Everyone talks about it
- It is implemented by engineers with some heavy math background
Looking at this very subjective summary, optimization hits maybe half of the items on the list. It’s not self-learning, not data-heavy, not spoken about in flashy newspaper articles.
So 1-0 for the non-AI crowd!
On the other hand optimization can, in some cases, fully automate a previously manual process. Take for example taxi trip merging, workforce planning or meal courier task assignment. Optimization is prescriptive data analytics and augments (and sometimes replaces) human decision making, which sounds like a very defining feature of an AI system.
Looks like the game is tied again at 1-1
Yet another angle is to contrast optimization with machine learning, perhaps the most familiar category of AI methods. Machine learning ticks most of the boxes in my AI feature list and if something currently existing can be called AI, it’s machine learning. The main differences between optimization and machine learning are:
- In machine learning the machine learns and produces insights; In optimization the business logic comes from humans.
- Machine learning models are inherently black boxes (although the insights can be explained, see e.g. https://christophm.github.io/interpretable-ml-book/). Optimization models are inherently transparent, i.e. it’s easy to track back to first principles on why the model outputs a certain result.
- Machine learning is predictive analytics; optimization is prescriptive analytics (suggesting decisions)
So where did we end up? Language is imprecise and my “AI” is not the same as your “AI”.
To me it may mean self-learning systems, whereas to you it may mean automating decision-making processes.
Nothing unusual in that – to a biologist the tomato is a fruit but my kids prefer their fruit salad without tomatoes. It doesn’t really help that we need contrived expressions like “mathematical optimization” and “integer optimization” to even get across what exactly we mean by optimization.
Maybe AI-optimization would be a good compromise or we could even use terminology like “optimization is one of the prescriptive analytics AI software technologies to enable better decision-making processes”, but that’s practically speaking quite long.
If you have good ideas on what to call it, do let us know! In the meantime we’ll continue helping companies do more with less.