Swarm Intelligence: Learning from Ants

Cars that drive by themselves and robots operating autonomously are an increasing feature of modern life. This is made possible by using state-of-the-art computer technology that simulates thinking and decision making processes in our brain. To prevent the many new road users from getting in each other’s way, they also need to learn cooperation. Researchers promise that traffic jams will be a thing of the past thanks to collective or swarm intelligence.

Back in the 1970s everyone in the IT scene was talking enthusiastically about artificial intelligence (AI). Autonomous robots would assist humans by performing dirty, dangerous or just downright unpopular jobs. But the excitement actually began in the United States in July 1956 during a six-week conference at Dartmouth College in Hanover, New Hampshire. Revolutionary ideas were up for discussion there. Computers could make intelligent decisions, it was claimed, and not just execute a rigid, unimaginative program.

They could learn by themselves and even draw conclusions from the results, faster and more precise than humans. Outside university campuses, many people would have regarded such ideas as insane. Most people in the USA and Europe only knew about electronic data processing, as it was called at the time, from head-in-the-clouds press reports. Marvin Minsky and Hans Moravec were the champions of this doctrine, constantly banging the drum to promote the subject among the general public. Almost a little too well. The hype was soon followed by sobering reality. As time went by, hardly anyone believed this prediction would come true just 60 years later. Not much happened after that. Researchers tackling the subject were confronted with the accusation of flogging a dead horse. Even so, small groups quietly set about working in the field of artificial intelligence.

Computers could learn by themselves and even draw conclusions

_{Marvin Minsky, AI pioneer}

According to their successors, there are many reasons why these pioneers failed at the time. Primarily, their knowledge of how the human brain functions was still very rudimentary, but the ins and outs of the original have to be understood to be able to replicate it technologically.

It also became clear that AI called for different programming languages and computer architectures to those for information systems commonly used for the financial world. The massive computers of the 1960s were fed with punched cards and cooled with water, the microprocessors currently installed in every car had not made it past the drawing board at that time. Today, manufacturers such as graphics specialist Nvidia produce processors with 15 billion transistor functions – and at prices to match the budgets of ordinary people. In Minsky and Moravec’s time, such a computer would have blown even NASA’s astronomical budget. Processors for graphics applications are especially suitable for AI as they are able to map a neural network very efficiently. These networks form the artificial brain with which a computer ‘thinks’ – a three-dimensional structure comprising diverse networked elements. In effect, a three-dimensional image. This makes manufacturers of graphics cards, which provide enormous computing power for relatively little money, heroes among IT pioneers. This is because neural networks need a high level of processing power plus a fairly long training time. Very similar to the brain of a child, which first has to gain experiences before it can reliably make correct decisions. With the power available in modern chips, it’s no wonder the manufacturers of graphics cards have discovered a lucrative new business opportunity. At the beginning of 2016, Nvidia CEO Jen-Hsun Huang announced a special computer for AI and driverless cars at CES in Las Vegas. The Drive PX2 has two powerful Tegra CPUs and two Pascal GPUs, which are among the most powerful of their kind on the world market. The four chips are said to provide as much processing power as a 150 MacBook Pros. This supercomputer for the car of the future uses 250 watts and is cooled with water. Though the power drain should not be a problem for larger cars and buses, a version for small cars would need to be more economical.

A swarm of underwater robots together find the shallowest point in a basin after collective comparison with another not-quiteso-shallow area.

Progress in semiconductor technology should solve this problem within a tech cycle or two. The enticing market for millions of cars that travel to and from on our roads has driven Nvidia to redouble its efforts to find suitable partners in the field of AI. The company has made a timely connection in Europe by incorporating the DFKI (German Research Centre for Artifcial Intelligence) into its GPU Research Centers scheme. Crucial here was the center’s expertise in the area of ‘deep learning’, the efficient training of neural networks with large data volumes. To teach a neural network the concept of a ‘house’, it would need many photos of lots of houses viewed from different perspectives. This used to be a challenge because gathering information cost a great deal of money. Nowadays, such a thing is child’s play thanks to the internet and the social media proliferating there. Today, deep learning’s main cost is time – pictures and examples are plentiful. Once a neural network has been trained, it automatically starts generating its own interrelations and tests the results. Deep learning is already used to search data downloaded from technical equipment and machines to determine early indicators of potential failures. That means maintenance intervals can be set more intelligently for individual machines.

Search engines are probably the bestknown application field. Without AI, it would not be possible to categorize and catalog the enormous amounts of data on the internet. It’s hardly surprising, therefore, to find the biggest search engine developer Google is also involved with the DFKI. In October 2015, the company became the 17th industrial partner of the DFKI. The degree of investment hasn’t been disclosed, but it’s said that Google may have invested as much as all the other 16 industrial partners put together.

Ant-inspired robots: with ‘scent trails’ emulated by phosphorescent floor dye.

It is not only the implementations that have changed dramatically. The theoretical aspect has also been worked on intensively. There are several distinct approaches to AI, each based on very specific concepts. Some involve the application of conventional principles of logic, while others tend to be inspired by statistical mathematics.

Advocates of the statistical approach are currently in the majority worldwide. The main question becomes: Is it the aim of AI to simulate the human mind as precisely as possible, or is it enough to use machine learning to make the best decision quickly? This is a question that demands a pragmatic answer from developers. Applications using AI are increasingly becoming reality, along with driverless cars and robots. But when numerous autonomous devices are operating together, they should ideally achieve more together than the sum of their individual actions. The blueprint here is nature. When bees and ants go about their work in swarms, waiting times are kept to a minimum. In contrast to road users, these living organisms neither intentionally nor unintentionally obstruct one another, even though hundreds are making their way to and from the hive or nest. This should also be possible for electronic drones, packed with pizzas or parcels, or for robots to achieve – or so the researchers hope. Perhaps this is why ‘intelligent’ cars have become the most important research application of all – because who enjoys sitting in a traffic jam?

A solution for this requires a lot more than precise sensors and a powerful neural network, a fact that quickly became clear to researchers. Bees and ants ‘talk’ to one another and some form of communication also needs to occur in the robot world. This is the central prerequisite for mutual networking and coordination. Work is already in progress on new data formats that enable this kind of communication and these protocols are already becoming a technical reality. Governments have also reacted and legal standards for vehicle-vehicle (V2V) communications will appear in the USA and Europe.

Swarm Intelligence: Seeking a common goal

How will this translate into practice? Cars will not tell each other where there are ample sources of food. Bees and ants do this. Cars will tend to ‘communicate’ about traffic congestion, or warn against problems. A heavily laden truck will know when it is approaching a hill, thanks to its digital road map, allowing its computer to niftily calculate that its journey will be slowed down. A simple signal to car users in the immediate vicinity will advise them to overtake and the truck will help to make this possible. Traffic lights will also receive and send signals electronically, thereby allowing their neural networks to adapt to a developing situation, even if the signal cannot be registered visually. While logistical systems like these are being implemented, it still lacks a swarm ‘mentality’. Organisms with swarm intelligence are characterized by their ability to perform actions as if they were an integral whole with a common goal to be served. Individuals striving for advantages and the desire for personal success are alien to swarm intelligence. The majority of traffic problems today are caused, or exacerbated, by impatient or aggressive drivers who force their fellow road users to brake abruptly, or they may even cause accidents. Professor Thomas Schmickl, from the Artificial Life Laboratory at the University of Graz in Austria, is one of the few scientists whose research involves robots and living creatures. He is inspired by the communication systems of fish, fireflies, honey bees, cockroaches and slime molds.

We seek our inspiration from fish, honey bees, cockroaches, and slime molds.

_{Thomas Schmickl, University of Graz}

In contrast to the DFKI, Schmickl’s department sets less store by intensive learning, focusing instead on the power of self-regulation. The autonomous underwater robots and small cars he uses achieve ‘smart’ final configurations more or less automatically. The learning curve required can’t be passed successfully, or safely, in real traffic but can be overcome in the lab. As part of the Collective Cognitive Robots (CoCoRo) project, funded by the EU, Schmickl is researching how autonomous units can fnd their way around like ants. Phosphorescent dye on the ground serves as a substitute for the scent trails ants use in nature.

Heterogeneous swarm of underwater robots heading out on a search task (magnetic field emitter).

At the DFKI, researchers are also avoiding the exclusive use of the deep-learning approach. If incidents only occur rarely, machine learning utilizing mass data is of little benefit. A simple example is a driverless car at a carnival. How should it respond if it suddenly encounters an elephant? Flee in panic? Swerve out of the way? Humans are superior to technology here, as they are continuously aware of space and time and can relate this to their current situation. “Artificial intelligence is still a tender blossom, but it’s set to mature rapidly, making it a powerful instrument,” says Doctor Joseph Reger, CTO of Fujitsu Germany. He has a vision of a combination of IoT, smart cities and AI. “The world is becoming a universal sensor”, explains Reger, “because AI gains new and previously concealed information independently from existing sensors and cameras. Most of this is already technically feasible today.” AI is still in its infancy. Hardly any research area is so interdisciplinary. Psychology, neurology and neurosciences, mathematics and logic, communication science, philosophy and linguistics come together – not forgetting the engineers who then have to put the parts together.

Telling a driverless car when to brake

The DFKI is the world’s largest research center for artificial intelligence. At sites in Bremen, Saarbrücken and Kaiserslautern, scientists study both the principles and the practical application of AI – and have been doing so for almost 30 years. If we encounter autonomous vehicles on the road or service robots vacuuming our carpets, this is in no small part thanks to what DFKI researchers have achieved. When the institute was founded in 1988, there was already a great deal of optimism that these predictions would come true. When a driverless car brakes today, it makes this decision in a millisecond. In the 80s, the calculations required would have taken an hour even in a high-performance computer. And those computers could not be mobilized – while the AI systems in today’s driverless cars are the size of handbags.

READ MORE ARTICLES