In his excellent novel „Do androids dream of electric sheep“ (the movie “Blade Runner” is based on this story) the famous science fiction novelist Philip K. Dick deals with the blurred line between artificial intelligence and humans. In Dick’s dystopian world, androids have become highly sophisticated to the point where they can hardly be distinguished from humans. A recent event that took place in a hotel in the Korean capital Seoul was also basically about the similarities between humans and sophisticated AIs. Here, the program AlphaGo, developed by the Alphabet Company DeepMind, competed with a human grand master of the ancient game Go. This game, which originated in China and is not quite as popular outside of Asia, involves a game board consisting of 19 horizontal and 19 vertical lines with 361 intersections, where usually two players alternately place white and black stones. The goal of the game is to encircle the stones of the opponent and thus take them off the game board.
More possibilities than atoms in the universe
So far the game was considered too complex for a machine to beat a human player. However, this particular complexity is not created by the game’s rules, which are quite simple. In the case of Go it is rather the sheer number of possibilities how a game can evolve from start to end. Other than, for example chess, which has a comparatively limited number of possible board configurations (on average 35 different decisions per round), Go has about 250 different options per move. To calculate all possible scenarios until the end of the game with a brute force method is next to impossible: A game of Go can assume 10*10170 different board configurations, which is far more that the supposed number of atoms in the known universe („only“ about 1089). Because of this fact, even experts placed their bets on the human player. The more or less surprising outcome: AlphaGo defeated the South Korean Lee Sedol in a best of five mode with 4:1.
This match stands in line with famous competitions between computers and humans: In 1997 IBM’s super computer Deep Blue defeated chess world champion Garry Kasparov. And in 2011 another IBM computer named Watson bested two human quiz champions in a game of Jeopardy.
AlphaGo relies on neural networks
However, one detail is very different to those past machine victories: While Deep Blue and Watson were trained by programmers to beat their human opponents, AlphaGo learned Go from scratch. Initially it studied about 30 million game positions and moves of human players from a database. Then it played several million games against itself, and after that, also played against other Go-programs (with a victory quota of 99.8%). The machine mimicked this special autonomous learning behavior from its creators.
The algorithms of AlphaGo rely on the principle of the Monte Carlo Tree Search and so called neural networks – information structures, which are modelled after the human nervous system. Artificial intelligence based on neural networks thus learns very similar to a human and could outperform them in this ability in the near future. Alpha Go consists of two different neural networks: One which proposes sensible moves and another which evaluates those moves by forecasting the probable outcome of the overall game.
Neural networks in Operations Research
It is quite obvious that neural networks basically can be used for areas of application other than games. In the end, the game Go is about finding the optimal decision for the next move out of a vast number of possibilities. Alphabet wants to use the findings in the AlphaGo project for e.g. autonomous driving, medical research or climate forecasts. This leads to the question: How are or will neural networks be used in Operations Research, for example as a basis for forecasting techniques?
Most of the forecast algorithms which are currently used for decisions in the area of supply chain planning, production planning or fraud detection are based on rather classical, statistical methods. Although neural networks are not really new to operations research – more than 5000 scientific papers have been published in the last 20 years – there has not been a real breakthrough. The results of an algorithm contest (the NN3 competition) published in 2011 showed that the classical methods stood their ground against the neural networks.
For some use cases, other established methods like Fuzzy Logic seem to be more suited. Neural networks have to learn and thus need recurrent input. In case of fraud detection, a human could very quickly formulate a rule for a new fraud pattern in a program with Fuzzy Logic, which becomes instantly active. A program based on neural networks needs to observe this particular pattern several times until it is able to identify it and implement the new rule. Until then the damage could already be done.
Conclusion
With its victory over Lee Sedol, AlphaGo has set a milestone in the history of artificial intelligence and has shown us what machines are capable right now. Although neural networks do not seem to play a dominant role in Operations Research today, it might be just a question of time until they catch up to the classical statistical methods. Until some days ago, no one would have imagined a machine beating a grandmaster of Go.
Maybe someday forecast algorithms will dream of electric sheep, who knows?
