OR Practice with 19th Century Optimization Technology

So you are an enthusiastic optimization guru with a MS/PhD in Industrial Eng/Operations Research. You want to bring your ideas to life in the private industry (Recession notwithstanding). You are the OR guy in the IT team who shuns heuristic approaches ever since that day your customer observed a major improvement in the objective function after adding a highly restrictive constraint, and wondered what kind of an idiotic optimization product (yes, he air quotes that) you were hawking. Red-faced, you decide to build a math programming based Mixed Integer Programming-based (MIP) solver using the cool 21st century stuff that OR folks in academia swim in. It's a NP-hard problem, but your ideas works great in practice and you get voted the 'employee of the month'. What's more, it satisfies all the sacred OR practice requirements of Rosenthal and Brown.

You may have had to shell out 15,000$ to buy a new CPLEX license (and that's just for development, buddy), or even better, a new Gurobi Dev license that apparently costs only half as much with no extra charge for parallel stuff. Now comes the challenge of putting all this into a product. Your Manager says "I don't care about optimality. The data is full of noise anyway". You feel sheepish, but you are a true OR believer. You realize that its not the optimal solution that matters, but the consistency in the response that is achieved by always finding (near-)optimal solutions. Your manager is now convinced but yours is a small company with great ideas, and the royalty costs for the MIP solver kill the profit margin. Your director tells you to come back with a better idea. You are shocked. You never had to worry about a solver back in school. It was always there in the optimization lab, after all.

You decide to go back to 20th century technology and work with just Linear Programs (LP). You somehow figure out a way to reconcile the fractional people and broken equipment to get a solution that features fully-limbed personnel manning machines with all working parts. Its still better than randomized heuristics, right? You find out that in the current economic climate even LP solvers are expensive. You begin to realize that this yet another reason why OR hasn't taken off in a big way beyond the niche markets. You love CPLEX, GUROBI, and other tools and the guys who built them. But you also learn that even though every decision problem in practice has constraints, only the large companies with a prior OR history tend to adopt the cutting-edge O.R required to robustly handle such problems.

Finally, after 18 months of back-breaking research and Dev, you realize that the randomized heuristic tool built in a day by the computer-science Dev guy is still in place. Folks begin to think the O.R works really well only in academia and theory. You dont give up. Instead, you decide to turn to 19th century technology. You contemplate building your own crude LP solver. Its time to reinvent the wheel ... to be continued ...