Self-Play Deep Learning and Poker
Problem being addressed
Game theory is often limited to studying games in restricted domains because the number of variables and parameters gets out of hand. How can deep learning be used to model and understand games? And develop strategy?
A deep learning strategy in which an AI trains by playing against itself is implemented. The main application is to large scale games of imperfect information that are usually too difficult to model from first principles in game theory. The AI learns through repeated self-play how to compute approximate Nash equilibria solutions to the games of Leduc Poker and Limit Texas Hold'em. It is found that over a long enough train period Neural Fictitious Self Play (their name for the strategy) is able to achieve superhuman performance in large scale poker tournaments.
Advantages of this solution
This is an interesting application of deep learning to game theory and to card playing. Since game theory is such an integral part of modern economics and bargaining this could have repercussions in the economic arena.
Solution originally applied in these industries
Possible New Application of the Work
Aerospace & Defence Sector
Simulations are an important consideration for the defence sector. This includes understanding the behaviour of other parties during conflict. This research could be reapplied to large scale defence problems in war hotspots.
Game theory is often used in business to model negotiations and other high risk situations where the lack of cooperation can lead to suboptimal outcomes for the business. This research could be reapplied to modelling mergers and acquisitions, for example, in order to better understand optimal bargaining strategies.
Travel and Tourism Industry
The online travel industry is extremely competitive. Deep learning approaches such as this one could be used to find strategies that are optimal for a given player in the travel and tourism industry, when other players' strategies are taken into account.
Source URL: #############