Optimization is one of the core technologies for Decision Management and an increasingly important one as the technology grows to handle more operational problems. Dynadec is one of the new players in this space and has what it regards as a game-changing optimization platform (Comet). Their intent is not simply to provide a new optimization tool but to drive value for large companies in vehicle routing, workforce management and resource scheduling. Keeping to these focus areas allows them to build more complete solutions, though the underlying technology has other potential uses. Dynadec sees an opportunity (and I agree with them) because the economic environment means that companies are trying to drive out costs and better utilize assets. Add in the massive growth in RFID/sensors/GPS and real-time data and there is an opportunity to use this data to solve some complex problems. Dynadec believes they can help companies make decisions they could not make before – quickly and under uncertainty.
Dynadec’s Comet supports what can be thought of as the whole spectrum of optimization – from mathematical programming, to constraint programming, constraint based local search and now what they call “dynamic stochastic combinatorial optimization.” This last is unique to them and combines predictive models and optimization so it can adapt model-based on predictions of the future. Comet supports these four kinds of optimization and allows access to/from the standard enterprise platform elements – databases for historical data, SOA, C++ etc. Built on this platform are the three solutions for vehicle routing, workforce management and resource scheduling. Dynadec provides hosted or on-site solutions as well as embedding the engine into ERP systems or providing it as a black box service. As time passes they are increasingly packaging these applications.
Today they have two broad categories of customers. Some use a proof of value pilot approach where Dynadec spend time working with them to show how it works. This is aimed at companies without the optimization expertise to use the platform and the end result is that Dynadec hands over a solution. Others work on an application development pilot where the customer’s optimization staff is trained on Comet and Dynadec focuses on skills transfer and helping with integration. These customers end up licensing the technology as a development platform.
Several demos followed:
- Vehicle routing – their example was for an oil company where people move from shore to platforms, between platforms and back to shore. Objective is to minimize the number of helicopters and it is complex because the helicopters are close to their range limit, must therefore be refueled, and people on multiple rigs must be scheduled. Dynadec did the pilot in about 3-4 weeks and the schedule was getting calculated in about 20 minutes – much faster – while reducing the amount of helicopter time required.
- Hot Strip Mill scheduling was the second one. Change over time was a problem and the mill had lots of constraints – wear, oxidation, grouping etc. This was an experienced optimization user that had hit the wall on their use of the existing technology. Using Dynadec they were able to satisfy all their constraints (something their existing solution could not achieve), to improve urgency and transition costs significantly. All of this was done in under two minutes.
- Rostering. Another sophisticated user of mathematical programming and other optimization technology. Hospital scheduling is tricky with different skills in the staff pool and complexity in assigning them to different roles in different time slots. This hospital had too many constraints and was just trying to find a solution. It took Dynadec 45 seconds to solve the problem more effectively.
- Last example is a company that needs to make a decision every 2 minutes around customers in a region who need to be visited. They have no advance knowledge of where customers are, when they will need service or what they will need. Essentially customers get added to the list of those who must be visited when they call in. The company can accept customers or reject them when they call but once accepted they must be served. This is an exercise both in continuous optimization – after every call – and in the integration of predictive modeling. Dynadec improves the rate of coverage by integrating their optimization with a model predicting the likelihood of a particular region having a customer at a particular time. Afterwards they have all the data and can compare how the engine would have optimized with perfect data with what it actually recommended. In other words you can assess how well you would have done if you had perfect advanced knowledge. Dynadec is using some machine learning algorithms to automate this feedback loop.
Dynadec’s technology is targeted at an interesting subset of the optimization / operational decisioning market. If the problem is too small then customers don’t care about optimizing it (or people can do the optimization). If the problem is very large then customers don’t care because the problem “grays out” – the average works well enough all the time. In between is the sweet spot – lots of data but not so much that blurs from one day to another. As an example Saturday delivery for a delivery company would be in the sweet spot while a weekday might not be. Air taxis might be another example or restoring the power grid after problems by dispatching skilled techs.
One of the features I most liked about Dynadec’s solution is that it attempts to be robust in the face of uncertainty by integrating predictive analytics and optimization. Their ability to get to a good answer (if not the best) quickly and the potential for using the solution in new classes of operational decisions becomes clear. This is critical as optimization technology typically iterates multiple times looking for the best answer. In operational decisions, though, you want the best answer you can get quickly so an optimization engine that gets most of the way to the answer early in its iterations is more useful. For example in one large scale vehicle routing problem – 90 vehicles and a thousand customers – they could find an optimal solution in an hour but could drop the total travel time almost to the optimal level after just 5 minutes. All this means that even high-volume operational decisions can be optimized one at a time.