Warning: This is kind of an opinionated article, based on the little experience I have. The world is Bayesian, and I may update my beliefs as I come across more intriguing problems in the near future.
One of my professors at Jadavpur University always used to emphasize on “connecting the dots”. He used to say to become an asset to an organisation one needs to master this art. Over the last 16 months of working, heuristics and empirics have been two of my biggest ‘friends’ when it boiled down to connecting the dots.
When I first started working, I was given a variant of a Facility Location Problem. And my manager said something that time: “Try coming up with your own heuristics. They are more flexible.” At first, I found it hard to understand, but over time I realised how right he was.
When it comes to mathematical modelling (and if you’re using an open-source solver), there are a few trade-offs that you need to take care of. If the problem size is huge, and there are too many real-life constraints, then you’re better off using your own heuristics than aiming for an exact solution. I remember one time I wrote a MILP, it almost took 2 hours for the MILP formulation and for the solver to solve it. Then I thought of changing the method, it took me almost 3 days to come up with my own heuristic. The result? The solution time went down to ~100 secs. Yes! Not even kidding! Albeit, the solution cost was higher than the exact solution, but if one wanted to put the whole thing into production then the heuristic would have been the far better choice. I am planning to write an academic paper on this so not going into the details of the method right now.
Next, I remember I had data on the Pincode level for different Indian cities. I had to convert them to latitude and longitude points so that every point would be around the given Pincode, approximately. This has to be one of the most interesting data pre-processing I have done so far. So what did I do? I used numerical methods! More precisely: I formulated the problem through an approximate equation and then used the Newton-Raphson method (learned this in my 3rd sem of Engineering, never thought would come in handy like this). Bam! Solved!
Now, as I had said earlier it is about those trade-offs. So, there will be situations when building a solution based on your heuristic will not be worth it. For me, so far in 15+ months, it had happened once. I had to solve a knapsack problem in R. The Knapsack function in R was erratic, and writing my own logic for that didn’t seem worth it as it was just a sub-problem of a bigger problem. So, I just formulated an LP and solved it. The solution time was very less and the results were satisfactory!
Nowadays, the term “connecting the dots” seems very pertinent to me for these reasons. And two of the best tools to do that have been: heuristics and empirics. The only downside is, that the art of connecting the dots isn’t taught proactively. It is something that you need to acquire. And yes, one more important thing which I strongly feel: before anything, aligning oneself with the business is the most crucial thing. If you’re not doing that in a religious manner, then it will be impossible to take a little help from these two friends!
2 thoughts on “Connecting the dots with heuristics and empirics”
Reblogged this on ORdinary Analysis.
Inspiring. No matter how long it takes, original solutions are often better.
LikeLiked by 1 person