Scientists at DeepMind have trained an AI to allocate resources effectively in an online game. The results of the study were published in Nature.
To create the algorithm, the researchers conducted an experiment using an economic online game in which participants determine how to distribute their resources for mutual gain.
The simulation involves four people, each of whom receives different amounts of money. Players must decide whether to keep it for themselves or contribute to a public fund that provides returns on investments. The return on investments can be adjusted so that some players benefit others.
Possible return-on-investment mechanisms include three models:
- a strictly egalitarian model, in which the incomes from public funds are distributed evenly regardless of contribution;
- a libertarian model, where payments are proportional to contributions;
- a liberal-egalitarian model, where each player’s payments are proportional to the share of their own funds that they contribute.
They asked groups of people to play many rounds under different levels of inequality and using various profit-distribution mechanisms. They then selected the preferred method.
Using these data, the researchers taught AI to imitate human behaviour in the game, including the voting stage. They pitted agents in thousands of rounds, while another algorithm tuned the redistribution mechanism based on votes.
In the end the AI settled on a model that most closely resembled liberal-egalitarian. However the algorithm returned almost nothing to players unless they contributed roughly half of their capital.
They then contrasted the AI method with the other three in a new stage of the game with humans. In a blind vote, the algorithmic method generally received the most votes, the report says.
According to the researchers, the AI mechanism likely performed well because payments based on relative contributions help correct the initial wealth imbalance. However, mandating a minimum contribution does not allow less wealthy players to benefit from the contributions of richer players.
DeepMind researchers stressed that their work is not “an instruction manual for building an AI government.” They also do not plan to develop AI-based tools for policymaking.
In May DeepMind presented a visual language model with 80 billion parameters.
In February the AI lab developed the AlphaCode tool, which can write code autonomously.
In August 2021, DeepMind introduced a universal architecture for building artificial intelligence.
Subscribe to ForkLog news on Telegram: ForkLog AI — all the news from the world of AI!