Yesterday I published a brief overview about the main types of games we can find in game theory and their relevance in deep learning or artificial intelligence(AI) scenario. Today, I would like to bring your attention to a new method that was recently published by Alphabet’s subsidiary DeepMind and that provides a unique way to tackle asymmetric game problems. DeepMind’s breakthrough can have profound implications in modern multi-agent, AI systems that are often modeled as asymmetric games. Before getting there, let’s try to understand what’s so difficult about asymmetric game environments.
Where Nash Fell Short: The Challenges of Asymmetric Games
Symmetric games are the favorite examples used to illustrate the dynamics of game theory as they are mathematically elegant and near perfect. As we explained in yesterday’s article, a symmetric game describes a dynamic in which the different players share the same strategy and goals. Typically, the simplicity of symmetric games makes it easier to model from a computational standpoint. Unfortunately, most real life game environments lack the mathematical elegance of symmetric games.
Asymmetric games describe multi-agent environments in which players have different and often conflicting goals and strategies. Let’s take yesterday’s market collapse as an example. In that environment, some traders were desperately trying to offload their positions while others were trying to accumulate new positions planning for a potential bounce back of the market (doesn’t seem is going to happen today based on the futures ;)). Multiply that strategy for the millions of traders and investors around the world and you have an incredibly chaotic asymmetric game.
In game theory, the solution to many asymmetric game environments is modeled using the Nash equilibrium. The model was named after John Forbes Nash, the American mathematician immortalized by Russell Crow in the movie “A Wonderful Mind”. Essentially, a Nash equilibrium describes a situation in which each player has chosen a strategy and no player can benefit by changing strategies while the other players keep theirs unchanged.
The Nash equilibrium is a beautiful and incredibly powerful mathematical model to tackle many game theory problems but it also falls short in many asymmetric game environments. For starters, the Nash method assumes that players have infinite computing power which is rarely the case in real world environments. Also many Nash equilibrium models fail to account for the notion of risk which is omnipresent in most asymmetric games the economic markets. As a result, there are many asymmetric game scenarios that are hard to implement using the Nash equilibrium. This is particularly important in multi-agent AI systems that need to find the right balance between the mathematical elegance of the solution and the practicality of its implementation.
Deep Mind’s Symmetric Decomposition of Asymmetric Games
In a recent paper published by DeepMind, the authors proposed a very clever model to find solutions for highly complex asymmetric games by decomposing them into different symmetric games. In mathematical terms, the new techniques proposes that if (x,y) is a Nash equilibrium of an asymmetric game (A,B), this implies that y is a Nash equilibrium of the symmetric counterpart game determined by payoff table A, and x is a Nash equilibrium of the symmetric counterpart game determined by payoff table B.
To illustrate the new technique, I am going to borrow an example from the original post in the DeepMind’s website. The example is based on the famous “Battle of the Sexes” game. Here, two players have to coordinate a night out to either the opera or the movies. One of the players has a slight preference for the opera and one of them has a slight preference for the movies. The game is asymmetric because, while both players have access to the same options, the corresponding rewards for each are different based on the players preferences. In order to maintain their friendship — or equilibrium — the players should choose the same activity (hence the zero payoff for separate activities).
This game has three equilibria: (i) both players deciding to go to the opera, (ii) both deciding to go to the movies, and (iii) a final, mixed option, where each player will opt for their preferred option three fifths of the time. This last option, which is said to be “unstable”, can be rapidly uncovered using DeepMind’s new method by simplifying — or decomposing — the asymmetric game into its symmetric counterparts. These counterpart games essentially considers the reward table of each player as a separate symmetric 2-player game with equilibrium points that coincide with the original asymmetric game.
Source: Deep Learning on Medium