Machines that play games against humans

This essay originally appeared in:
Edinburgh University Science Magazine
Issue 14, p. 19, 2013 (link)

Suppose you are playing the following game: You are in an internet chat room with two other participants. One of them is a human and the other is a machine. You are allowed to chat to each of them individually, but they can only chat to you and neither of them can see your conversation with the other one. In the end, you win the game if you can tell with certainty which one the machine is, or else the machine wins.

This is a variant of a game proposed in 1950 by Alan Turing, today considered one of the founding fathers of artificial intelligence. He proposed this game to address the question of whether machines can be intelligent and whether or not they can “think”. There has been much debate as to whether Turing's game truly addresses this question. However, it also sparked a vision of a future in which intelligent machines are ubiquitous parts of everyday life and where humans interact with machines naturally and effortlessly, just as they do with other humans.

How does playing games help us realise this vision? The fact is that interaction between humans and machines can be an extremely complex matter. Games allow us to study different aspects of interaction systematically and in isolation by providing well-defined environments for interactive decision making. The hope is that understanding these various aspects will eventually enable us to realise innovative applications, such as fully autonomous robots used in child care and nursing homes.

From a technical perspective, what makes games interesting is the fact that what we should do depends crucially on what the other players are doing and how they react to our moves. A machine that is to be successful in this setting must be able to plan its actions in the face of uncertainty and learn from experience. This becomes particularly difficult if the other players are human, since human behaviour can be very complex and is often a priori unknown.

The Robust Autonomy and Decisions (RAD) group of the University of Edinburgh studies human-machine interaction in several ways. One line of work is concerned with the “ad hoc coordination” problem. Therein, the goal is to design an autonomous decision making agent which can achieve flexible and efficient interaction with other agents whose behaviours are initially unknown. This research has led to a novel agent, called HBA, which compares the other players' observed actions to a set of familiar behaviours in order to find optimal responses.

In 2012, the RAD group organised a robot exhibit for the Royal Society Summer Science Exhibition in London. Part of the exhibit was a human-machine experiment in which over 400 participants played repeated “Rock-Paper-Scissors” against HBA and other agents. The purpose of the experiment was to show that a small set of behaviours could account for a much larger set of complex human behaviours. One interesting result was that HBA won significantly more games than the humans. This was despite the fact that in theory the optimal strategy was for both players to play randomly.

While many of the games studied in artificial intelligence do not explicitly model physical interaction, there are also games in which physical interaction is a central component. One example is given by the annual RoboCup football competition, in which teams of robotic football players engage in an adapted version of the real football game. The ambitious aim is that by the mid-21st century, a team of autonomous humanoid robots will be able to win a match against the winner of the most recent (human) World Cup.

The RAD group is the first and currently only UK-based research group to enter a team, called “Edinferno”, in the standard platform league of the RoboCup football competition. The team made it into the quarter finals of the 2012 competition in Mexico and will compete again in the 2013 competition in the Netherlands.

There is no conclusive answer as to whether machines will (or should) ever pass the Turing test. However, the past decades have seen major progress in human-machine interaction, and researchers will continue to push the boundaries.