SEOUL, SOUTH KOREA - MARCH 13: In this handout image provided by Google, South Korean professional Go player Lee Se-Dol (R) prepares for his fourth match against Google's artificial intelligence program, AlphaGo, during the Google DeepMind Challenge Match on March 13, 2016 in Seoul, South Korea. Lee Se-dol played a five-game match against a computer program developed by a Google, AlphaGo. (Photo by Google via Getty Images)
Go player Lee Se-dol, right, prepares for battle against AlphaGo in the 2016 Google DeepMind Challenge. Google’s AI program was victorious © Google/Getty

Be the first to know about every new Coronavirus story

We are going to need a lot of ingenuity to exit from our pandemic-induced economic slump. Can powerful machine learning systems, similar to Google DeepMind’s game-playing AlphaGo, help us out of the mess?

That is certainly the hope of a group of researchers from the software group Salesforce and Harvard University, who last week unveiled the AI Economist, one of the most innovative attempts to apply machine learning to economic modelling.

The possibilities for this fast-developing research field, known as algorithmic economics, are intriguing. Some of the discipline’s techniques are already being used in narrow ways by technology companies, such as Google and Amazon, to model auctions and set prices. But the aim of the AI Economist is to help empower policymakers to develop and test proposals in a dynamic, simulated environment. 

As one of the most outspoken champions of stakeholder capitalism, Marc Benioff, chairman of Salesforce, has long trumpeted the social purpose of business. Salesforce has therefore developed the AI Economist with the explicit goal of improving social welfare as well as optimising economic performance. 

The first focus of its research has been tax policy and the trade-offs between productivity and equality. Longer term, the research team believes the AI Economist could shed light on many other areas, such as the impact of stimulus measures, or universal basic income, or environmental regulations.

In essence, AI Economist is a data-driven computer simulation that uses reinforcement learning techniques to model millions of interactions between different actors in the economy, such as workers and governments. Rational agent programs, trained on historic data sets and acting in a defined economic environment, are given an objective function and then set loose to achieve their goal any way they can. As with AlphaGo, when it defeated the champion Go player Lee Se-dol in 2016, such programs can generate startlingly novel insights.

In the case of the AI Economist, the model suggests a different tax schedule from those most commonly used around the world. It reckons the optimal design is to set higher tax rates for the rich and lower tax rates for middle-income earners, generating additional wealth to subsidise those on low incomes.

Richard Socher, Salesforce’s chief scientist, highlights the AI Economist’s current limitations but argues that the model can become an increasingly sophisticated tool for policymakers. He likens the effect to sequencing the human genome: the model dramatically gains in usefulness once scaled up. 

For the moment, the AI Economist only uses four AI agents in a Minecraft-style environment and does not account for many real-world factors, such as inherited wealth, voluntary work or the ability to hire employees or set up a company. The interactions of an economy are also massively more complex than any board game. 

“Our simulations are not accurate or realistic enough at this point,” Mr Socher concedes. “AI agents are hyper rational. But people are not always super rational. Jamaica is happier than Germany even if the objective measures would suggest otherwise.”

Editor’s note

The Financial Times is making key coronavirus coverage free to read to help everyone stay informed. 

Find the latest here.

To improve the model, the research team is keen to engage with other computer scientists and economists to enrich its “minimum viable economic system”.

David Parkes, a Harvard professor who researches algorithmic economics and helped develop the AI Economist, says one of the strengths of the model is that it can use historic data to understand better how people are likely to act in practice. 

“We can use real world data to try to infer what utility functions people have. What do they optimise for? How does it change over time?” he says.

The AI Economist team is conscious, though, that existing data sets do not fully reflect marginalised members of society. They are also wary of politicians misusing such models to advance partisan arguments.

Tony Yates, a former Bank of England economist, says the AI Economist is an incredibly stylised economic model but it is in some respects a lot richer than many others currently used by central banks. 

“Machine learning is really just another form of econometrics, of pattern recognition,” he says. “It is fascinating. But I am confident that you cannot yet read anything from it in the real world as regards to what actual tax policy should be.”

Economists need to embrace diversity and complexity and engage with other disciplines to deepen their understanding of how the world works. Algorithmic economics is a tantalising, if faltering, step in that direction. 

The discipline may yet live up to the statisticians’ incantation: all models are wrong, but some are useful. 

john.thornhill@ft.com

Follow John Thornhill with myFT

Get alerts on Artificial intelligence when a new story is published

Copyright The Financial Times Limited 2020. All rights reserved.
Reuse this content (opens in new window)

Follow the topics in this article