DeepEcon is an AI-based toolbox for analyzing dynamic economic models with a large number of heterogeneous agents. Each agent solves a second order stochastic difference equation, and the individual solutions are restricted to satisfy certain aggregate equilibrium conditions. The solution is to be given in the form of the Markov decision process. The economy’s state space is the entire distribution of the individual quantities, so that the model suffers from a severe curse of dimensionality. This class of problems is analyzed in economics by approximating the state space with a finite set of aggregate moments, however, such approximations have important limitations.
The novel feature of the DeepEcon.org toolbox is that it ameliorates the curse of dimensionality by using the-state-of-the-art techniques from modern data science. First, it uses deep neural networks to perform model reduction by condensing information from many input state variables into a smaller set of features in hidden layers. Second, deep neural networks are robust to multicollinearity which helps me deal with ill conditioning. To reduce the cost of function evaluations, the training is performed using stochastic optimization. To construct expectation functions, asymptotically unbiased Monte Carlo integration is used. Finally, the toolbox makes use of parallel computing on GPUs to achieve further reduction in cost. Taken together, these techniques enable DeepEcon.org to solve models with thousands of state variables.
The toolbox is designed in a simple and intuitive way that enables users to easily apply it for solving their own models and applications. A user just updates the model parameters, neural networks, transition equations and runs the code. The toolbox is written in TensorFlow and PyTorch – the Google and Facebook data platforms that facilitate most remarkable data science applications such as image and speech recognition, self-driving cars, etc.