Artificial intelligence as structural estimation: Deep blue, bonanza, and alphago
By: Igami, Mitsuru
.
Material type: 
BookPublisher: The Econometrics Journal Description: 23(3), Sep, 2020: p.S1-S24.Subject(s): Approximate dynamic programming, Artificial intelligence, Conditional choice probability, Deep neural network, Dynamic structural model, Inverse reinforcement learning, Optimal control, Reinforcement learning, Simulation estimator
In:
The Econometrics JournalSummary: This article clarifies the connections between certain algorithms to develop artificial intelligence (AI) and the econometrics of dynamic structural models, with concrete examples of three 'game AIs'. Chess-playing Deep Blue is a calibrated value function, whereas shogi-playing Bonanza is an estimated value function via Rust’s nested fixed-point (NFXP) method. AlphaGo’s 'supervised-learning policy network' is a deep-neural-network implementation of the conditional-choice-probability (CCP) estimation reminiscent of Hotz and Miller's first step; the construction of its 'reinforcement-learning value network' is analogous to their conditional choice simulation (CCS). I then explain the similarities and differences between AI-related methods and structural estimation more generally, and suggest areas of potential cross-fertilization. - Reproduced
| Item type | Current location | Call number | Vol info | Status | Date due | Barcode |
|---|---|---|---|---|---|---|
Articles
|
Indian Institute of Public Administration | 23(3), Sep, 2020: p.S1-S24 | Available | AR124352 |
This article clarifies the connections between certain algorithms to develop artificial intelligence (AI) and the econometrics of dynamic structural models, with concrete examples of three 'game AIs'. Chess-playing Deep Blue is a calibrated value function, whereas shogi-playing Bonanza is an estimated value function via Rust’s nested fixed-point (NFXP) method. AlphaGo’s 'supervised-learning policy network' is a deep-neural-network implementation of the conditional-choice-probability (CCP) estimation reminiscent of Hotz and Miller's first step; the construction of its 'reinforcement-learning value network' is analogous to their conditional choice simulation (CCS). I then explain the similarities and differences between AI-related methods and structural estimation more generally, and suggest areas of potential cross-fertilization. - Reproduced
There are no comments for this item.