11 Jan Six things you need to know about evolutionary computation, with Professor Emma Hart
Evolutionary computation – robots learning from their environment, thinking about what they need to complete a task and redesigning themselves to do it – might seem like the stuff of science fiction, but this is the end goal of Professor Emma Hart’s latest project.
In partnership with three universities in Bristol, York and Amsterdam, her team is using a field of study called evolutionary computation, which takes inspiration from nature to create processes that have applications ranging from factory logistics to composing music.
Here are six things she teaches us about evolutionary computation:
1. Evolutionary computation reduces tasks that could take until the end of time to a matter of minutes
“We don’t normally look at evolution as an optimisation process, but evolution is about searching for the right combination of genes that an organism should have to survive,” Emma says. In the same way evolution efficiently searches through billions of potential combinations of genes, evolutionary algorithms are able to use similar principles inside a computer, much faster than manually testing all possible combinations with mathematical formulas.
2. We don’t yet know why these algorithms make their choices
Due to the nature of the process, it’s difficult to know how or why these algorithms choose their solutions. “We can have a method of deciding how good a solution is. For example, if it’s planning a route between multiple places, you can judge how good it is based on how long the route takes, the cost, the CO2 emissions etc. But, we don’t know what processes it went through to get there,” Emma explains. This also means that ethics have to be carefully considered by humans, as an algorithm may try to achieve its goal at any cost. “Say if you’re trying to maximise profit, it might do it but at the expense of people, so you have to be very careful how you define quality.”
3. It has many applications, from the creative to the functional
Evolutionary computation is mostly used in logistics, scheduling or business processes like creating staff rotas. However, these processes have also been used to design websites and wind turbines, more efficient gas pumps and even antennas on spacecrafts that were deployed by NASA. “We see them being used to evolve art, paintings and music. There’re some really interesting design applications,” Emma says.
4. The algorithms can be even more clever than people
Because the algorithms are not constrained by the way humans are taught to design, they can put forward really innovative solutions. In the example of the gas pump, the algorithm created a design that Emma describes as “something that didn’t look like anything any engineer would have ever dreamt of designing”.
This comes with a price though; because it can be creative, these algorithms can exploit bugs in code that humans might never have known existed. “There’s an example where software engineers tried to use these processes to fix programmes. They gave it some test cases and scored the algorithm based on how many it got wrong – the fewer the better,” Emma says. Within minutes, the algorithm was showing that it was getting none of the test cases wrong. “But when the team looked more closely, they realised the programme had learnt to delete all the test cases – so there were no cases to get wrong, therefore it got them all right.”
5. Evolutionary computation is not actual intelligence
“If you look at it dispassionately, evolutionary computing is basically doing a fast search of all the possibilities and rating them accordingly. It’s not intelligence in the sense that it understands the problem it’s trying to solve,” Emma says. “I wouldn’t use the word intelligence.”
That said, recent innovation in the field has led to the algorithms “thinking more like a person”, according to Emma. “For many years, the way evolution in a computer worked was driven by quality, so we were driving the evolution to find the best quality solution possible. But recently there’s been a focus on searching for novelty or diversity rather than just quality. Rather than searching for the measure that you want, by searching for interesting or novel ideas you tend to stumble upon the best solution quicker.”
6. It could change the way we think about machine learning
The field has come a long way from its early days, but the algorithms still have limitations. While they work well in theory, applying them in real life situations is challenging as the algorithms are not yet good at adapting if anything in the situation changes, like if a delivery arrives late or there aren’t enough people on the day to do a task. “This means you then have to retrain the algorithm or, better still, the algorithm should learn by itself to adapt to the new situation,” Emma says. “It’s a big shift in the way we think about AI. Machines shouldn’t just learn to do a job well, they should continue to learn and improve with experience.”
Emma’s most recent project involves pairing evolutionary algorithms with a 3D printer, allowing them to print robots that then explore a situation and ‘reproduce’ by printing off new models that are better adapted to it. “We’re looking at how to decommission a nuclear reactor, where you can’t know what’s inside. So, robots are redesigning themselves based on the challenges that they encounter. You can imagine one day using this process in space or deep-sea exploration,” she says .