In 2030 AI and big data driven Robots will be as normal in our daily routines and business environments as the internet today. I love the automotive industry as it is the most obvious example of that upcoming new normal: an autonomous car is a robot on four wheels, driven by a highly skilled AI and fed by big data. Very soon we will be talking about automobiles as opposed to human driven cars that will be as exceptional as a horse in the city.

The challenges that developers, companies and governments are facing when it comes to automous vehicles reveal the issues we will see in every industry and at a larger scale in society.

Train the brain

How can you create an AI? By making it smart. There are 3 ways to do that.

Optimization/planning. That is to model and plan effects of actions and plan ahead what actions to take. In other words: know the game and plan all potential scenarios. Think chess. Or Go. Artificial Intelligence is so much better here than human beings. Traffic however is very different from a board-game. Planned intelligence is something we may be able to use on a closed circuit but is pretty inaccurate when talking about releasing a car into everyday traffic, which is far to random.

Reinforced learning. Trial-and-error-based learning over time. An ideal way to learn to play a computer game, given that the loops are long enough to learn and numerous and that there are enough impulses to actually be able to do many iterations of trial and error. But this method is not a good idea when it comes to  training a vehicle in real life as traffic is not really a laboratory, but a situation in which real lives are at risk.

Imitation Learning. Collect human demos and try to imitate. When I drive my Tesla, even when I have not enabled my autopilot, it is observing me as a shadow pilot all the time, comparing how it would react to my actions and the actions of 100.000s of drivers all over the globe. We have been feeding the central brain with billions of kilometres of data, so that one day, that brain will outperform us and is ready to take over.

A lot safer

AI run vehicles are ready to take over. They are far better and more trustworthy than human beings. In the US only, people kill 40.000 people in traffic every year. On average all of us kill another person in traffic every 1.8 million kilometre. So the very moment a robot does a fraction better, we save 1000s of lives. If all cars would be AI powered, the number of fatal accidents would be close to zero.

The problem with autonomous vehicles are not the rules, not the sensors not the computing capacity nor their experience. The interaction with us, human creatures makes it extra challenging. Algorithms have to learn that in between what they know is best and performing that, there are humans and humans are pretty unpredictable. Just imagine autonomous vehicles merging (*) in between human driven ones or crossing a road without traffic lights. Either the robot is too polite and immobilized, or it is to aggressive and causes accidents anyhow. When all cars are autonomous, the problem would be solved. Not in a hybrid world.

Training robots to understand that humans are not a barrier

As a result, in between that fully autonomous Day After Tomorrow and Today, there is a Tomorrow that is most challenging. In the future robots will have to interact with humans, and not consider us a barrier in between them and their goal. Computers need to learn to interact. The most difficult element to overcome in this learning curve is the fact that computers need to ‘realise’ that their actions can influence human behaviour. Algorithms have to understand their own behaviour and ‘read’ the effects it has on humans to really interact. In order to make sure computers understand us, we have to be able to explain why we act the way we act and why certain actions are ‘good’ and others are ‘bad’ and that good and bad are not absolute but related to circumstances. It forces us to deep-dive into essential human issues that we never really had to explain to other humans, but will be essential when talking to those alien brains.

The lessons learned are these: in an AI world the biggest challenges will be the training of the AI in a safe environment and the interaction with human beings. Training will always involve big data and the interaction with humans will make us question ethics, sociology, psychology, morality and even philosophical issues.

(* In traffic engineering, the late merge or zipper method is a convention for merging traffic into a reduced number of lanes. Drivers in merging lanes are expected to use both lanes to advance to the lane reduction point and merge at that location, alternating turns.)

Rik Vera Please keep your mobile device
in a upright position (portrait mode).