multiagent-reinforcement-learning

I proposed a machine learning approach to model cognition and biases in investors’. Presented model can be conceptually divides into – Agent States and Learning. The Levels of rationality or Agent States are modeled through Multilayer Feed-Forward Neural Network trained with Error Backpropagation algorithm. The model was trained with historical data of listed stocks from the Bombay Stock Exchange (B.S.E). Model representing least irrationality achieved a Root Mean Square Error of 0.017 and co-relation of 0.99 and 0.98 in the train and test set respectively. Various networks were trained to different values of RMSE to obtain different states. A reinforcement learning (Q-learning) approach was presented for investors’ learn- ing. Belief Persistence, Availability Bias and Information Bias were modeled using various parameters of the Learning algorithm and introduction of a concept of irrational update. A psychological game and Bayesian Network was presented for probabilistic estimation