Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Social Environment Design

Edwin Zhang, Sadie Zhao, Tonghan Wang, Safwan Hossain, Henry Gasztowtt, Stephan Zheng, David C. Parkes, Milind Tambe, Yiling Chen

TL;DR

Social Environment Design is introduced, a general framework for the use of AI for automated policy-making that connects with the Reinforcement Learning, EconCS, and Computational Social Choice communities, and gives a direction for the systematic analysis of government and economic policy through AI simulation.

Abstract

Artificial Intelligence (AI) holds promise as a technology that can be used to improve government and economic policy-making. This paper proposes a new research agenda towards this end by introducing Social Environment Design, a general framework for the use of AI for automated policy-making that connects with the Reinforcement Learning, EconCS, and Computational Social Choice communities. The framework seeks to capture general economic environments, includes voting on policy objectives, and gives a direction for the systematic analysis of government and economic policy through AI simulation. We highlight key open problems for future research in AI-based policy-making. By solving these challenges, we hope to achieve various social welfare objectives, thereby promoting more ethical and responsible decision making.

Social Environment Design

TL;DR

Social Environment Design is introduced, a general framework for the use of AI for automated policy-making that connects with the Reinforcement Learning, EconCS, and Computational Social Choice communities, and gives a direction for the systematic analysis of government and economic policy through AI simulation.

Abstract

Artificial Intelligence (AI) holds promise as a technology that can be used to improve government and economic policy-making. This paper proposes a new research agenda towards this end by introducing Social Environment Design, a general framework for the use of AI for automated policy-making that connects with the Reinforcement Learning, EconCS, and Computational Social Choice communities. The framework seeks to capture general economic environments, includes voting on policy objectives, and gives a direction for the systematic analysis of government and economic policy through AI simulation. We highlight key open problems for future research in AI-based policy-making. By solving these challenges, we hope to achieve various social welfare objectives, thereby promoting more ethical and responsible decision making.
Paper Structure (14 sections, 9 equations, 2 figures, 1 table)

This paper contains 14 sections, 9 equations, 2 figures, 1 table.

Table of Contents

  1. Introduction
  2. Preliminaries
  3. Formal Definition of Social Environment Design Game
  4. Social Welfare Examples.
  5. Example: Apple Picking Game
  6. Challenges and Open Problems
  7. Related Work
  8. Economic Policy Design and Simulation
  9. Stackelberg Game
  10. Multi-Agent Reinforcement Learning
  11. Computational Social Choice
  12. Automated Mechanism Design
  13. Conclusion
  14. Environment Hyperparameters and Training Details

Figures (2)

  • Figure 1: The proposed framework. The process begins with voting, where human or AI players report preferences on social welfare objectives to a voting mechanism (1). This an objective for a Principal policy-maker, who designs a parameterized $N$-player Partially Observable Markov Game (POMG) (2). The players are the same as the voters. This POMG unfolds over several timesteps $H$ (3). Following the POMG, game state information is extracted to initiate $n$ last POMG state used as the first game state of the new round. This whole process is repeated again in the next round.
  • Figure 2: An example of Social Environment Design as an apple picking game, built with Melting Pot 2.0 agapiou2022melting. Player agents observe a restricted view of their environment, and receive a mixed reward depending on the apples they collect and the apples their observable neighbors collect. The principal observes both an unrestricted view of the environment and the running totals of all the players’ cumulative rewards, where it collects tax and redistributes wealth on the cumulated rewards at the end of every tax period (50 timesteps), similar to zheng2020ai. Rewards for the Principal are determined by the change in value of the objective it is currently assigned by agent votes. For training details and hyperparameters, please refer to \ref{['apenndix:hp']}.

Theorems & Definitions (9)

  • Definition 2.1
  • Definition 2.2
  • Definition 2.3
  • Definition 2.4
  • Definition 2.5
  • Definition 3.1
  • Definition 3.2
  • Definition 3.3
  • Definition 3.4