Game Theory and Autonomous Vehicles
Research on autonomous vehicles (I will refer to them as “autos”) has seen success, with companies such as Google and Tesla already having self-driving vehicles. While autos can no doubt efficiently travel from one location to another in a static environment, interactions with pedestrians and other road users can pose a threat. If the driver or pedestrian understands that an auto is expected to halt for anyone, then it is easy to see how humans can take advantage of this. Human drivers, of course, would never let this happen, and are able to successfully communicate and predict how other drivers and pedestrians will act. Research needs to be done, then, on how to simulate this behavior, so as to prevent the autos from continually stopping for anyone and everyone, preventing it from reaching its destination. This is especially important if autos are ever to be used for public transportation.
The paper discusses the “sequential chicken” model as a means to simulate two road users meeting at an intersection (one car at south, the other at east). This is a more complex version of the standard game theory model of Chicken. In Chicken, there are two players, both drivers. Each driver can choose between two strategies: swerve to avoid the other car, or continue straight. The player that swerves is considered the loser of the game (-1 payoff for this driver, +1 for the other), but of course if both continue straight and collide the payoff is much worse for both drivers (-100). If both drivers swerve than the game is a considered a draw (0,0). If you make a table to represent the payoffs for each choice, we can see that there is no pure Nash equilibrium for this game. In this game, only one choice is made and the outcome is immediately decided. In sequential chicken, multiple choices are made, where the two drivers continually approach the intersection at a speed they choose, until they finally have to decide whether they should swerve or not. The methods for determining payoff and the like are a bit more complicated, though, so I won’t discuss them here. Nevertheless, participants were asked to play in this sequential model game in hopes of determining how autos should react in similar situations. Participants were told that there main is objective was to avoid a collision and cross the intersection as short of a time as possible. In one session of games, the players were given chocolates as well, depending on whether they won (two pieces), lost (one piece), or crashed (zero pieces). The results showed that the chocolate games had more collisions, as well having less time delay (i.e. the players arrived at the intersection faster).
The paper ends by saying that although the model is by no means perfectly representative of what could happen in real-life situations, it is still a stepping as stone for further research. I agree with this; similar to how local properties in graphs can help teach us about global ones, so can simplified versions of games teach us about more complex versions of them. Additionally, we see another way game theory applies to real life. Even the simplest of interactions, such as crossing the street when a car is nearby or negotiating with another driver at an intersection can be modeled by game theory. In many cases, there is not a pure Nash equilibrium, and instead a mixed Nash Equilibrium is present. Game theory will no doubt play a role in artificial intelligence and the like, as the world around us continually become more automated, and thus requires human interaction.
Source: http://eprints.whiterose.ac.uk/129303/1/Camara2018MB_Manchester.pdf