Neural Networks in Artificial Intelligence & Game Theory for Deep Learning
In recent years, there have been vast improvements in artificial intelligence, specifically aided using neural networks to achieve more human-like capabilities. The artificial neural networks mimic real biological neural networks as the nodes of information are connected in a directed network, with sending and receiving signals. Where a biological brain sends messages to the cells within a body, a computerized neural network takes the incoming information, usually a set of large data, where it then goes through another set of nodes. However, the processes in these nodes are known as hidden layers, as this is the stage in computing where we are uncertain as to what happens (pictured below). Using machine learning techniques, the computer arrives at an output, which further extends the neural network and imitates the firing of a biological neural network.
Advances in artificial intelligence are further helped by networks called adversarial networks. These networks use advanced game theory techniques to deep learn information, making the computer able to teach itself many techniques and strategies for computing, games, etc. This type of environment is enabled by imperfect information games, which is different than the kind we have studied in class. With deep learning, the players (computers) are able to form and play strategies based on only your personal information and without knowing what other players know. When presented with an imperfect information game, the computers are still able to form a Nash Equilibrium and best responses to other strategies. Originally, adversarial networks used a method called fictitious play, its capabilities limited to computing the average strategy, which converges to the Nash Equilibrium of the game. However, this technique is limited as it cannot adapt to its own best strategies.
The most recently adopted technique to play these imperfect games uses the neural networks. Neural fictitious self-play is described as when each player consists of their own neural network with the ability to identify strategies, reinforce, and partake in supervised learning, instead of solely relying on computation. Each time the neural network goes through a data set, it plays a game according to game theory. Depending on the outcome of that game, the machine learns about the outcome and teaches itself to do better based on the probabilities, strategies, and previous instances of the game and other players. This is an increasingly complicated way to deep learn, but at the same time this is creating massive steps toward artificial intelligence and its ability to discern strategies at the rate of human comprehensive learning. These advances rely heavily on the game theory techniques learned in class but go even further to predict more real-life situations than very controlled prisoner’s dilemma environments. The use of directed neural networks and game theory are paving the way for revolutionary artificial intelligence.
https://www.kdnuggets.com/2016/10/artificial-intelligence-deep-learning-neural-networks-explained.html
https://arxiv.org/pdf/1603.01121.pdf
https://medium.com/intuitionmachine/game-theory-maps-the-future-of-deep-learning-21e193b0e33a