Artificial General Intelligence is not just one among many current research concepts, but a small dream of mankind. With Deep Reinforcement Learning, this dream seems within reach. The article shows where we are today and what road still lies ahead of us on the way to Artificial General Intelligence.

Subsidiaries of tech giants like DeepMind from Google or OpenAi from Tesla and SpaceX have realised numerous research projects in the field of reinforcement learning in recent years. This research is intended to Artificial intelligence generally better understood and, in the long term, the overarching goal of Artificial General Intelligence be achieved.

What is Artificial General Intelligence?

At Artificial General Intelligence is understood as the research project to create an artificial intelligence that possesses an intelligence comparable to that of humans. More precisely, this means that this Artificial General Intelligence is able to do so, all possible tasks to do. Today, intelligent programmes or Algorithms able to take on only individual, in some cases highly specialised tasks.

Link tip: If you are interested in concrete application examples of artificial intelligence, read our blog article on the topic "Deep Learning in practice".

The beginnings of DeepMind: solving Atari games

The official corporate goal - the creation of Artificial General Intelligence - of Google's DeepMind was considered pure utopia just a few years ago. Today, AIs already solve tasks at a level that exceeds the abilities of humans. What has so far only been treated in science fiction is nowadays already being realised by means of Reinforcement Learning implemented.

Google recognised the opportunities of reinforcement learning and AI very early on. As early as 2013, Google invested about 365 million euros in DeepMind, which at the time was a research institution that was working on solving Atari games. According to Google, DeepMind managed to develop both the rules of the game and success tactics on its own.

Reinforcement Learning heralded the turning point in AI research

DeepMind's AI in many games reached the capabilities at the level of Professional players and often even surpassed them. At this point at the latest, it was clear that the potential of Deep Reinforcement Learning was enormous. The DeepMind research project thus marks the turning point since increased attention has been paid to this method and its possibilities.

The challenge: One of the most difficult games in the world - Chinese Go

The ancient Chinese Go game is one of the most popular games in the world. It was considered one of the most Complexity as impossible for computer programmes to learn, let alone win. It is estimated that on the 19×19 board there are more than 10¹⁰⁴⁸ games possible. For comparison: in chess there are an estimated 10¹²⁰ possible games.

The sheer number of possible game variants becomes even clearer when one realises that - according to current findings - the estimated number of protons in the observable universe is approx. 10⁸⁰ is. In addition to the number of possible games, there is another difficulty with Go. Not every possible move can be explained on the basis of a Rules logic. Time and again, the masters of the game emphasise that their Intuition is the decisive element of the game.

Link tip: In our blog article on Machine Learning we explain why companies should care about an AI winning in the casino.

AlphaGo - an intelligent programme achieved the impossible. Today AlphaGo Zero takes its place

What has become known are the achievements of Google's artificial intelligence DeepMind probably primarily through the work on Alpha Go, the first AI based on DeepMind. It achieved what was previously thought impossible: AlphaGo defeated the world's best human Go players with aplomb. AlphaGo Zero, the latest sequel to AlphaGo, is even able to play the game by itself, without human intervention to learn.

AlphaGo Zero succeeds in this because it is based on Reinforcement Learning based. The AI begins to play completely randomly. Only the initial position of the pieces is given. After AlphaGo Zero had trained for three days, the AI competed against the first version of AlphaGo. The result: the Zero version was able to defeat its predecessor 100:0.

The next frontier: DeepMind AlphaStar

What's the latest challenge now that an artificial intelligence can even master one of the world's most difficult games better than humans? The answer is: Real-time strategy games. In January 2019, a new milestone was reached in this context. The latest AI, AlphaStar, defeated the elite e-sportsmen "TLO" and "MaNa" from the team "Liquid" in the strategy game Starcraft 2.

In comparison to Go, Starcraft 2 is a game in which in Real-time is played - in contrast to Go, where it is a turn-based game and there is plenty of time for each decision on each move. Rather, it is a complex strategy game in which it is a matter of correctly assessing the opponent and developing long-term strategies. Complicating the game is the fact that the AI does not have all the information at any time, as the radius of vision of the individual units is limited.

Deepmind AlphaStar (Source: https://techcrunch.com/2019/01/24/starcraft-ii-playing-ai-alphastar-takes-out-pros undefeated/?guccounter=1)

Conclusion: Are we on the verge of achieving Artificial General Intelligence?

The progress that has been made in the field of artificial intelligence research in recent years is groundbreaking and in the truest sense of the word revolutionary. No one - except the boldest dreamers and science fiction writers - would have foreseen a few years ago how quickly the Milestones are achievable. Moreover, the results far exceeded expectations every time.

Does this mean that we will be able to create Artificial General Intelligence in the next few years? Here go the Opinions are a matter of debate. One of the reasons for this is that researchers disagree on how exactly artificial general intelligence should be defined. What is clear is that intelligent algorithms, in particular thanks to Reinforcement Learning, are capable of solving more and more complex tasks. In many cases, the resulting programmes are already better than their human counterparts.

The most important thing is to apply the possibilities that have already been successfully experimented with in the game to real business problems. Based solely on the potential of reinforcement learning that actually already exists, the question of Artificial General Intelligence also plays a subordinate role. For the time being, it is important to use the already explore existing possibilities and useful areas of application to identify and develop them.