Last month AlphaGo, a self-learning AI machine, finally beat the Go world champion, Lee Sedol. As well as being a significant moment in Artificial Intelligence, this could also be a watershed for how we analyse environmental change and look for sustainable solutions to world problems.
Go is an ancient eastern game played on a 19 x 19 square board with two players who each take it in turn to place stones on the intersections of squares on the board. The idea is to encircle your opponent’s stones to take territory and capture their stones. The rules are simple but it has many possible outcomes due to its high branching factor and so is played primarily through intuition and feel, rather than rote learned moves.
To understand the high branching factor of Go it can be compared with chess. Chess has a board of 8x8 and each game lasts about 40 turns, a Go board is 19x19 squares and the average game length is 150 turns. This is why it has been very difficult to build a computer to beat world-class Go players who win though instinctive and experiential play.
AlphaGo is built using general-purpose AI methods and deep neural networks to mimic expert players. It can learn from past games (it’s watched 160,000 human-played, so called ‘supervised learning’) and self-learn by playing games against itself 24 hours a day for months on end. The effect of this isn’t to learn all possible positions, but to learn positions or combinations that’ll lead to a positive outcome.
So the amazing thing about the AlphaGo machine is that it can learn to base its decisions on the probability of a particular outcome, rather than a binary truth. A chess computer by comparison wins because it analyses every possible move before making a move. AlphaGo has developed a sense of intuition, something we associate with human and not artificial intelligence.
This level of machine learning I believe is a watershed for theorising, learning and most importantly solving environmental problems. Why?
Environmental problems are complicated, they have many interrelations and are influenced my many different elements, often global in nature. Environmental issues have a very high branching factor and they are not necessarily solved by trying to uncover the absolute truth to every variable.
A favourite example is weather predictions. The Met Office have built multi-million pound super computers which are often no more accurate than local fishermen or farmers who use experience and instinct to process information with better predictions. Would an AlphaGo computer, with some supervised learning (from correct farmer and Met Office forecasts), and followed by a period of self-learning have a better chance of predicting the weather than a traditional computer programme?
Could we better understand climate change if we added deep neural networks to our climate change computer modelling? Can we understand the likely impact and suffering caused by Zika if we build an AI machine to learn about malaria or the spread of Ebola in West Africa?
The high branching factor of environmental issues, whether they are the spread of a plume following a volcano, or the impact of CO2 emissions on global temperatures would benefit from AlphaGo’s approach to AI.
In addition to a new paradigm for learning and understanding, we would also see significant opportunity for sustainable environmental solutions. Autonomous vehicles already use very complex AI systems. Add in machine-learning and vehicles that never cause accidents, are much more fuel efficient and learn where they need to be at any given time, could be on our roads very soon.
However, imagine a world where your fridge and cooker/extraction fan had figured out that you want to make some pancakes from the data it had obtained; prompting a delivery that afternoon. 2 eggs, 300ml of milk and 250g of flour would arrive in your reused containers by an autonomous vehicle that’s already in your neighbourhood. It involves a lot of data with a high branching factor, but it might be possible soon.