This is an interesting and promising take on the problem. Despite being introduced already in the 60'ies, the optimization of delivery routes is still not used as widely as it should. I'd argue that this is mostly due to the complexities and challenges inherent in adapting such optimization technology to solve real world delivery route planning tasks, and, on the other hand, the high cost and low availability of operations researchers with relevant software engineering background.
In my recent PhD dissertation I tried to address the challenges from a different angle: I proposed using machine learning to predict the most suitable heuristic algorithm and its parameter values for a specific logistics planning problem. This way the developer or the user does not need to worry about the details of the optimization solver. The book is freely available for download from: https://jyx.jyu.fi/handle/123456789/65790
Hi Founders! As a developer, I would like to see 1) API documentation, 2) code examples 3) open-source version (whatever restricted in solutions quality). 1 and 2 can quickly give me some idea if this tool match my needs (and personal taste, to be fair). Open-source version is (often mandatory) escape hatch for many cases.
So far, I found nothing like this on your site. If your target audience are CTO/managers only, then it's OK, but developers would ask for real meat)))
How does this compare with open-source software like Optaplanner [0] from RedHat or ChocoSolver? [1]
I think this is really cool, and I've recently been looking into using constraints solvers for automating some of the manual scheduling processes at our startup and optimizing them (we do a variant of events booking, so essentially booking events at the times most likely to fill up and maximizing for profit)
[0] https://www.optaplanner.org/
OptaPlanner is an AI constraint solver. It optimizes planning and scheduling problems, such as the Vehicle Routing Problem, Employee Rostering, Maintenance Scheduling, Task Assignment, School Timetabling, Cloud Optimization, Conference Scheduling, Job Shop Scheduling, Bin Packing and many more. Every organization faces such challenges: assign a limited set of constrained resources (employees, assets, time and/or money) to provide products or services. OptaPlanner delivers more efficient plans, which reduce costs and improve service quality.
OptaPlanner is a lightweight, embeddable planning engine. It enables everyday Java™ programmers to solve optimization problems efficiently. It is also compatible with other JVM languages (such as Kotlin and Scala). Constraints apply on plain domain objects and can call existing code. There’s no need to input constraints as mathematical equations. Under the hood, OptaPlanner combines sophisticated AI optimization algorithms (such as Tabu Search, Simulated Annealing, Late Acceptance and other metaheuristics) with very efficient score calculation and other state-of-the-art constraint solving techniques.
Congratulations on the launch! I have a few questions.
What happens when the value function for a given state is expensive? The value is the distance from one state to another right, not the 'value of that state' - what if there are a lot of states that are "feasible neighbors"?
What sort of scale of routing problem are you guys taking? How many stops? Does DD scale well when you add in additional conditions like Driver Hours, Battery/Fuel Capacity?
I'd love to hear more about Dash. Is itself a generic simulation that people have to, effectively, program themselves? Or is it implementing supply chain-specific logic like ordering or fulfillment?
Cheers.
Congrats on the launch nextmv!
I had a chance to play with nextmv's beta when they first published it. By far the most useful aspect was the ability to bracket the decision results by calculation time. E.g., I can say, give me your best result of this choice given 100ms.
This changes the typical "train / test / deploy" ML process, to something where you can get as accurate a result as possible given some block of time. This gives you a lot more options when the value of having a super-precise decision drops off a lot after say, 80% accuracy.
For those of you familiar with rocketry, the technique is a lot similar to a Kalman runner [1]. Essentially when a rocket needs to gimbal adjust its trajectory, it has a ton of uncertainty about the nature of the environment, but it does an excellent job of making a fast educated guess for the simple purpose of "get me to this orbit and don't crash".
More generally, this gets to the core of the issues discussed in part 1 of the a16z article about AI companies [2]. Specifically that modeling to get to accurate result is a huge and hidden cost of ML, which makes it distinctly different from software startups. Decision science is an attempt to bridge that gap.
[1] https://www.bzarg.com/p/how-a-kalman-filter-works-in-picture...
[2] https://a16z.com/2020/02/16/the-new-business-of-ai-and-how-i...
Congrats on the launch!
I've worked with Ryan a couple times over the years since 2009. I also interviewed with Carolyn once upon a time as well. These founders are the brightest in the industry!
They have developer's experiences at the top of the priority list. That's not something I see when looking at existing tooling in this domain. The problem they are solving is real and applies to so many industries both obvious and not and this is a modern toolset to bring in to your stack.
I cannot wait to see the many successes they will enjoy. I encourage anyone with an optimization / constraint solver in their stack to take a look at Nextmv.
I think that is a beautiful toolset you have built. Really like the way DDs fit into modelling problem domains as state machines. I have signed up and will check it out.
Super cool and congrats on the launch! I love this idea, and if you are hiring, would consider working there just for the challenge alone. I've implemented the Held-Karp on the GPU in webgl if that counts for any bonus points :) I hope you all do well!
As far as I understand, this is for developers to develop algorithms, right? If that is the case, aren't the algorithms mature already? Or are we as developers just passing the data to existing algorithms and problem types of your system?
I'm an undergraduate who recently started entering the field of OR, and the idea of relaxing decision diagrams to make them polynomially sized sounds very interesting — the first time I hear about it.
Hey Carolyn and Ryan, how would you say this compares to the Google OR-Tools and their TSP solver? https://developers.google.com/optimization/routing/tsp
Congrats on the launch!
traveling salesman as a service? interesting
The topic you are tackling is a very hard one, but you have figured that one already. Good thing is, you have both a solid and relevant background.
A general thought, which might be important in the future. This kind of optimization tools work best at scale, I would go even as far as saying only at scale. And that can make a MVP a hard thing. I have seen people try something similar without the necessary scale, or rather density as routes were spread out over to large a region. And it didn't work out that good.
What I didn't get, but then might be just me, is who your customers are. Developers or users? If it's the latter I could provide some operational insight.