Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Optimization Solver as a Service

quicopt.com

Post reply on HN
Print view View original

41–48 of 48 posts

Re: Optimization Solver as a Service

#41

Post by DannyBee »

akoboldfrying wrote: post #21
DannyBee wrote: post #13

NEOS will let you run this stuff on cplex/gurobi/etc (IE much faster than the backends behind quicopt), for free, is integrated with pyomo/etc, and has like an 8 hour time limit. Often, the difference on "harder" problems is 10x or more. I have problems that gurobi solves in 30 seconds that take 15 minutes or more for ~every non-commercial solver (or-tools, HIGHS, ipopt, etc). But right now, this wouldn't even be int…

> But right now, this wouldn't even be interesting to me to use even if they actually were fronting commercial solvers, because they can't actually run it any faster So you use NEOS, but another service offering the same thing as NEOS would not be useful?

Not if i have to pay for it and use a different API?

Remember, NEOS is both free, and pyomo already supports it natively without me changing anything - i can use both neos and local without any issue.

Why would i move to something with a different API that i have to pay for?

Re: Optimization Solver as a Service

#42

Post by whatever1 »

ge0ffrey wrote: post #31
whatever1 wrote: post #14

We do have standard benchmarks in the field. Hans Mittelman maintains a library. No idea why they did not bother to run them. https://plato.asu.edu/guide.html Their website has just 3 cherry picked instances and claim complete dominance.

We run the Mittelman VRPLib benchmarks at Timefold (and beat other open source solvers like or-tools in 95%+ of the X datasets). But they are not representive of the real world, at all. The Mittelman VRPLib benchmarks have only 1-2 constraints. Skills? No need. Working hours? Unlimited. Maps integretion? Cars can fly and the earth is a flat Euclidean space. Any VRP algorithm optimized for the vrplib datasets is overf…

That is fair, which is why you should not over-index on academic benchmarks. But a frontier solver should be within a small margin from the competitors in these benchmarks. If you are 10x slower, there is no way you can catch up (at least not without a ton of work from the practitioner side).

Re: Optimization Solver as a Service

#43

Post by DannyBee »

MILP wrote: post #37
DannyBee wrote: post #13

NEOS will let you run this stuff on cplex/gurobi/etc (IE much faster than the backends behind quicopt), for free, is integrated with pyomo/etc, and has like an 8 hour time limit. Often, the difference on "harder" problems is 10x or more. I have problems that gurobi solves in 30 seconds that take 15 minutes or more for ~every non-commercial solver (or-tools, HIGHS, ipopt, etc). But right now, this wouldn't even be int…

Just curious, what kind of problems are you solving?

In this particular case they are power usage and rate optimization problems to do peak shaving/demand shifting with a combination of batteries and optionally, solar. It tells you the max you can save, the amount of battery/solar you should have (including the pareto front because sometimes it's like 2x the battery for $5 more in savings), and how to program the inverters.

For free, mind you, this is not part of a paid offering on my part.

These are easy for the case of non-demand rates (IE the rate just changes at x hour to x price), and can be solved by HIGHS/et al in a second or two. You can actually prove there is at least one optimal solution that only changes inverter programming at a rate change point.

They are actually quite complex when the rate has a demand charge (IE you are charged not just for x price per kwh, but also some amount * max demand usage of any single hour in a month).

The max demand charge is usually 80% of the bill.

The complexity is because recharging the batteries (particularly without solar) is the same as any other load from a demand perspective. So they have to be trickled (or charged from solar), etc. On top of that, lots of inverters have a limited amount of TOU slots you can use (for example, sol-ark inverters only support 6 periods). Which constrains it painfully. Gurobi can solve it in about 30 seconds. HIGHS takes around 15 minutes to solve it for 2 years of hourly history data.

Re: Optimization Solver as a Service

#44

Post by akoboldfrying »

DannyBee wrote: post #41
akoboldfrying wrote: post #21

Earlier quoted context omitted.

> But right now, this wouldn't even be interesting to me to use even if they actually were fronting commercial solvers, because they can't actually run it any faster So you use NEOS, but another service offering the same thing as NEOS would not be useful?

Not if i have to pay for it and use a different API? Remember, NEOS is both free, and pyomo already supports it natively without me changing anything - i can use both neos and local without any issue. Why would i move to something with a different API that i have to pay for?

> Not if i have to pay for it and use a different API?

TFA says Quicopt supports Pyomo. The example uses the "free tier", so you're not paying for anything (yet -- though I'm confused by the phrase "one-time entry point").

NEOS has generous limits, but the fact that Gurobi are still in business tells me that NEOS by itself can't satisfy everyone. Were Quicopt to offer access to Gurobi, etc., per your hypothetical, I think this would clearly be valuable.

More strongly I don't think it's crazy to offer "just" publicly available solvers, despite how much weaker they are than commercial ones. If I had to solve a stream of optimisation problems that were individually pretty easy but the rate of arrival was unpredictable, using such an "Optimisation as a Service" would make sense in much the same way that it makes more sense to serve spiky web traffic by spinning up cloud VMs on demand than by buying a bunch of on-prem boxes, even though those cloud VMs might be very weak.

Re: Optimization Solver as a Service

#45

Post by akoboldfrying »

deanalyzer wrote: post #19

This could be interesting, but it badly needs systematic benchmarking results. It is not difficult to get Claude Code or Codex to install and run a solver locally, so the tool’s current value proposition is fairly muddled. If there were evidence that it offered better performance, I might consider running larger workloads on it.

As I understand it, the value proposition is that it can deal with a spiky stream of problems to solve better than buying a bunch of on-prem hardware and running local solvers on them would -- similar to why it often makes sense to spin up cloud VMs on demand to handle spiky web traffic.

Performance on an individual problem is still interesting of course, but maybe not the main focus.

Re: Optimization Solver as a Service

#46

Post by luiwammus »

LPisGood wrote: post #39
luiwammus wrote: post #17

Earlier quoted context omitted.

The main reason why companies might prefer Hexaly is their emphasis on quickly finding high Quality feasible primal solutions.

I get that this is their goal, however they aren’t particularly strong in this area compared to free alternatives, traditional MIP, and custom heuristics.

I guess it depends on what your application is here. Hexaly primarily focuses on Routing, Packing and Scheduling problems, which have strong Combinatorial components. My experience with those type of problems is that Hexaly is in its own league above the MIP solvers, especially at scale. For applications like portfolio optimization or network design MIP solvers still fare a lot better.

Re: Optimization Solver as a Service

#47

Post by tirutiru »

throwaway832040 wrote: post #25
ge0ffrey wrote: post #23

Earlier quoted context omitted.

There is an audience for such platforms - Timefold Platform optimizes 1,000,000 visits and 2,000,000 shifts per week - but only if it's more than just orchestration. If it handles explainabily, what-if scenarios and insights to fulfill business needs. And that's where supporting many solvers becomes the blocker. A lowest common denominator design. Those solvers are a black box. They don't expose what they're running,…

None of these solvers genuinely focuses on the quality of the features that matter in real-world operations. Many of them, including Timefold, lack a realistic, financially grounded model of the world. They do not adequately account for traffic, driver preferences, or other factors that require a continuous feedback loop between what actually happened in practice and what the optimizer expected to happen. A vehicle-r…

I work on vehicle routing for a living. Our routing engine is mostly original heuristics, no MIP.

The thrust of your remarks is correct IMO. We are forced to support a large number of granular constraints, too large for my comfort.

But no solver will deal with large 'global events'. That's just not its job. Different levels of decision making altogether.

My email is in my profile. Would be happy to hear more.

Re: Optimization Solver as a Service

#48

Post by DannyBee »

akoboldfrying wrote: post #44
DannyBee wrote: post #41

Earlier quoted context omitted.

Not if i have to pay for it and use a different API? Remember, NEOS is both free, and pyomo already supports it natively without me changing anything - i can use both neos and local without any issue. Why would i move to something with a different API that i have to pay for?

> Not if i have to pay for it and use a different API? TFA says Quicopt supports Pyomo. The example uses the "free tier", so you're not paying for anything (yet -- though I'm confused by the phrase "one-time entry point"). NEOS has generous limits, but the fact that Gurobi are still in business tells me that NEOS by itself can't satisfy everyone. Were Quicopt to offer access to Gurobi, etc., per your hypothetical, I…

Sure. If it supported commercial solvers for me it would definitely be more valuable. I do think it’s fairly crazy to pay for remote solving using public solvers - the case you give is uncommon enough that whoever has it would probably want their own infrastructure to deal with it. Beyond that the current setup of trying to sell a single call solver as your advantage when the customers who might buy it for real need flexibility of various sorts doesn’t make sense to me.

FWIW-By supports pyomo they mean I can install their python package and hand a pyomo model to their solver API. I don’t want that. I want to use pyomo’s solver API which is how I interact with all other solvers

In the end I just don’t see who they are selling to. If their advantage is a single call solver for different model kinds, which is what they spend the most time talking about, that could be done with a small python package, and would already exist if it was a big issue

Post reply on HN