Unmanned Systems Technology 007 | UMEX 2016 report | Navya ARMA | Launch & recovery systems | AIE 225CS | AUVs | Electric motors | Lethal autonomous weapons

77 accidents in more than 2.4 million miles of autonomous and manual driving combined and, says Google, “not once was the self-driving car the cause of the accident”. The vast majority involved the self-driving car being rear-ended while stationary or at very low speed by cars driven by humans. None though has yet been faced with the kind of extreme situation in which its AI might have to choose between crashing into a queue of people waiting for a bus or into a bridge parapet, say, or any other situation in which it has to decide who to hit. There is a strong argument that, as there are no good outcomes for such situations – which are extremely rare even for human drivers – it is wrong to expect the AI to make the ethically ‘right’ decision. In short, the argument goes, the question of what the AI would do in the face of such an ethical dilemma is wrong. AI in general is improving rapidly, showing the ability to shine in more and more complex environments. Ever since IBM chess computer Deep Blue beat world champion Gary Kasparov in 1997, it has been clear that computers will beat the most capable humans at tasks that reward the rapid application of clear rules to situations that involve very large numbers of combinations and permutations. Many other games, including poker, draughts (checkers to US readers), computer games such as Space Invaders and even reverse quizzes such as Jeopardy! (in which contestants are given answers and must work out what the questions are) have fallen to AI. AI masters Go The latest of these is the ancient Asian board game of Go, the object of which is to place counters to surround a larger area of the board than one’s opponent by the end of the game. Go is played on a board of 19 x 19 squares to rules that, while simple, lead to extremely complex strategies. In 2000, German mathematician Achim Flammenkamp ran simulations revealing that there are almost 2.1 ×10 170 legal positions in Go, far in excess of the current estimate for the number of atoms in the observable universe, which is about 10 82 protons, or hydrogen atom equivalents. Google DeepMind, creator of the artificially intelligent AlphaGo program, says this makes the game a googol (1.0 x 10 100 ) times more complex than chess. On March 15, 2016, AlphaGo beat world champion Lee Sodol by four games to one in a match played over four days, the first time an AI program has defeated a top-ranked Go professional. That is significant because Go is said to have too many possible moves for a computer to win by brute- force calculation alone (constructing a search tree over all possible positions), which is how Deep Blue beat Gary Kasparov. AlphaGo’s creator Demis Hassabis and his colleagues took a different approach, combining an advanced tree search with deep neural networks. According to Google’s official blog, these networks take a description of the Go board as an input and process it through 12 network layers containing millions of neuron-like connections. A ‘policy network’ selects the moves to play, while a ‘value network’ predicts the winner. With neural nets trained on 30 million moves from games played by human experts, AlphaGo then played thousands of games between its neural Unmanned Systems Technology | April/May 2016 The accidents that Google’s self-driving cars have been involved in have been minor, with most of the fault having been attributed to other vehicles (Courtesy of Google) Many games including poker, draughts and Space Invaders, and even quizzes such as Jeopardy! have all fallen to AI