Optimization Solver as a Service
quicopt.com
quicopt.com
This doesn't solve or provide guidance for the subtle problems in these otherwise opensource solvers... The first example requires the client to manually disambiguate equivalent variables to get a stable solution... Sure that's a pretty common problem everyone working with optimizers should be familiar with but they're also one of the hardest things to track down in a complex derived model.
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, why they made certain decisions or how they can scale to large datasets or complex business requirements.
We've picked our poison: one solver, which we've built in the open, in the last 20 years, versatile enough to handle any scheduling problem. That delivers.
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-routing problem without real-world feedback is little more than a gimmick. Even assuming the world could be modelled perfectly, what happens when an unpredictable event disrupts the plan? Is the supposedly “globally optimal” solution robust enough to adapt, or will it create a backlog that forces the business to hire additional workers because the system failed to build in sufficient redundancy?
Using MILP makes the system even less flexible.
Neither our clusters in the US or EU can afford to go down for a minute, or business operations in logistics, retail and healthcare are impacted.
> They do not adequately account for traffic, driver preferences, or other factors that require a continuous feedback
- Traffic: supported
- Driver preferences: the APIs support
-- Area affinity (soft) and geo fencing (hard)
-- Fairness and load balancing of work (soft)
-- Overtime (soft and hard)
-- Seniors not doing boring junior tasks (soft)
- Continuous feedback: the APIs support
-- Real-time rescheduling (warm starts) as actual execution data comes in.
-- Recommend assignment: scheduling in survival mode
-- Pinning: user stays in control through overriding assignments
-- Explainability: why a certain decision was made
-- Insights: what are the bottlenecks in my schedule - what type of employees should I hire or upskill.
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.
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 interesting to me to use even if they actually were fronting commercial solvers, because they can't actually run it any faster and having this ".solve" API does nothing - pyomo already does that for me in practice.
So you use NEOS, but another service offering the same thing as NEOS would not be useful?
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?
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.
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
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.
That's our Solver as a Service for scheduling problems (vehicle routing problem, shift scheduling, job scheduling, etc). It runs scheduling problems implemented with our open source solver: solver.timefold.ai
But this post is such a service for formula problems instead (think master capacity planning, portfolio optimization, etc), due to the choice of MILP solvers underneath. Similar to NextMv, Neos, etc.
Misc thoughts:
- I'm not familiar with the LABS problem, but the LABS benchmark page is interesting & compares against Gurobi. I'd be curious to see how an existing commercial non-mip approximate solver such as Hexaly (formerly LocalSolver) compares here.
- the other two benchmarks aren't very convincing as they don't compare against other methods or show running times
- the front page mentions peer reviewed methodology - consider linking to the publications
- good idea to have case studies of applications. I was a bit confused to see this listed under 'References' but on comparison the Gurobi & Hexaly marketing websites also do this (references -> case studies & references -> customer stories, respectively)
- re the client API, you may want to make the server URL have a default, so your trial users / customers don't have to specify it. It may be easier for you to roll out changes to your server URL in future if you can do it by changing the default server URL in a new version of your client library rather than requiring your customers to update their source code.
All the best!
Hexaly claims to go far beyond MIP. Amazon uses it for packing VMs into servers. This video by one of their research scientists was widely circulated at the time [1].
I work on combinatorial optimization too but a specific problem so we write the heuristics from scratch. Seems exact solvers are doing a lot more these days?
I created a json like schema/struct/whatever to describe the problem. Maybe adopt something like this and more people will be able to see how they could use your tool:
https://github.com/JWally/jsLPSolver/blob/master/API.md
I need to re go through the docs, but you get the gist.
Here is the Berlin Airlift problem for example:
const model = { optimize: "capacity", opType: "max", constraints: { plane: { max: 44 }, person: { max: 512 }, cost: { max: 300000 }, }, variables: { brit: { capacity: 20000, plane: 1, person: 8, cost: 5000 }, yank: { capacity: 30000, plane: 1, person: 16, cost: 9000 }, }, };
https://plato.asu.edu/guide.html
Their website has just 3 cherry picked instances and claim complete dominance.
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 overfitted and not the best one in reality.
Take HGS for instance. Brilliant for CVRPTW. Crumbles to dust in field service routing for telco operations etc.
If there were evidence that it offered better performance, I might consider running larger workloads on it.
Performance on an individual problem is still interesting of course, but maybe not the main focus.
OR-tools is almost exclusively linear programming which according to its strict assumptions converges more or less trivially, assuming a correctly composed program.
Which means if you're paying for it "as a service" you all but deserve to lose that money.
> Different algorithms are better for different problems
So... why does your rhetorical style have such oppositional tone if you're just going to reaffirm the no free lunch theorem?
Some our customers want to go even higher, and we're working on that.