“We solve hard problems,” says the top line on Satalia’s website.
Based in London, they are a firm that offers optimisation services to a range of clients. Using artificial intelligence and machine learning techniques, they can help solve – or at least significantly improve responses – a variety of challenges, from staff scheduling to infrastructure planning to, that old favourite, the travelling salesman problem. This effort was recognised last year when Gartner put them on their Cool Vendors in Data Science list, where they were the only company from the UK.
Satalia’s growing war chest of algorithms contains huge potential for the retail sector, with its elaborate supply chains, massive resourcing demands and large staff numbers. Indeed, they already count a number of household-name British retailers as clients; we caught up with CEO Daniel Hulme to discuss the nature of what optimisation actually is, how it works, and where its future lies.
Can you offer a definition of optimisation in a single sentence?
At its most basic level, optimisation is process of trying to find the best solution [typically] to a mathematical problem in a reasonable amount of time. Most people don’t realise how insanely complex a task this is.
What can it offer that we humans can’t?
Machine learning is really a set of techniques that allows you to find patterns and make predictions from data. Humans are very good at finding patterns in three-dimensional worlds, like perceiving that something is a chair, or predicting the path of a moving object.
Machine learning, on the other hand, is really good at finding patterns in data. It’s particularly good at finding patterns in multi-dimensional spaces – we can do it in three, but machine learning techniques can find patterns in many, many dimensions, where there are many different variables.
However, they are not necessarily very good at making decisions – there’s a big distinction here between finding patterns and making decisions.
Decisions are where I imagine the idea of intelligence creeps in.
The best definition of intelligence I’ve ever found is ‘goal-directed adaptive behaviour.’ ‘Goal,’ as in you’re trying to achieve a goal, ‘behaviour,’ as in allocating my resources in a way that’s going to achieve that goal, and ‘adaptive’ in that I can continuously make myself better without any intervention. Machine learning techniques are very good at making predictions, but they’re not actually achieving any goal.
So goals are the key?
It’s not until you actually give them an objective – say if I want to predict if some customers are going to churn, what do I need to do? What I need to do is allocate my best sales person to make sure that that customer doesn’t churn. So machine learning is really about predicting things, and then optimisation is about satisfying some objective, and for me an AI system isn’t an AI system unless it’s finding patterns from data, making decisions to achieve a goal and then learning from those decisions. You very rarely see systems in production that actually learn from the decisions that they make.
Everyone gets excited about AI. I think there’s going to be a bit of a bubble because we’re expecting machine learning to solve all our problems, but actually these techniques are probably going to fall foul to the same biases that humans have. We are really good at finding patterns but we’re not very good at making decisions. Say I asked you to optimise some deliveries of some vehicles, dropping 20 items to 20 customers; it’s impossible for a human to come up with a perfect solution or even a very, very good solution is impossible for a human to do. It’s also going to be very difficult for any pattern recognition system to do, so you need optimisation algorithms to solve the problem.
So it acts as an alternative to brute force processing power on complex calculations?
Brute force for complex problems is impossible. The example I use is that if I’ve got 24 points on a map and I wanted to know what the shortest path is between those points, the brute force approach would take 20 billion years to come up with the best one – which clearly isn’t practical. So you have to use optimisation in the real world.
Can you give a real-world example?
Delivery routes are perhaps the most well-known case. Predicting how long it will take you to drive from A to B – this is where machine learning is good, and optimisation is taking that data and figuring out what to do with it. Any allocation of resources like this is an exponential problem – 24 points on a map doesn’t sound like much, but it gets incredibly complex very quickly. You also need all the data you can get on things like roads, traffic and the weather – even religious holidays, which obviously the coding can’t determine for itself.
It’s worth it, though – it can lead to pretty amazing efficiencies. Some of our clients are saving millions or tens of millions of pounds for very small marginal improvements. If you can shave a minute off every delivery, and you are making 100,000 deliveries in a day, there’s clearly huge savings to be made.
Staffing seems an area where optimisation could have a huge impact.
Absolutely – scale is obviously a factor here. We’ve been working with a major consultancy firm. Say we’re optimising seven thousand of their staff – saving 1 per cent of their time works out at five minutes a day. Extrapolate that out for that many people over even a year, and that tiny adjust can result in weeks’ worth of time saved.
There’s actually an overlap with delivery optimisation here: if staff are located at points where, say, they are crossing one another on their journeys to work, that’s both inefficient and not good for them, unless they love to commute. You want to use machine learning to identify which staff members are right for what jobs, which staff members work well in which teams, and then figure out how to best allocate them.
Could there be any cultural resistance from people to having their lives to some extent determined by software?
The company we’re working with thinks of it in terms of staff happiness, rather than pure cost savings. They are working with their staff to help them manage the change, interviewing their staff to make sure they understand and they are on board. I understand it can on the surface sound very new, let’s say, but there’s actually huge potential, if it’s done properly, to remove lots of red tape and bureaucracy. At Satalia, we take our ideas and implement them internally; we don’t have managers or administrators at all.
Where next for optimisation? What’s the big challenge to solve?
Modelling these problems can still be a big hurdle. We still need smart people to actually go into businesses and find out the rules within their system, like say legal limits to working hours, decisions on workplace culture and so on, and then model them mathematically. Once you give them to algorithms to solve, that’s actually the straightforward bit – there are a huge number of them out there.
No, the big problem is modeling the problems you give to the algorithms. I think there’s going to be a boom in the skills required for optimisation; there’s already a boom in data scientists, but there’s going to be a renaissance in operations research – people who are skilled at logic and so on. We will see tools appearing that allow people to model these problems better.
More information on what Satalia offer can be found here.