![]() ![]() Moreover, these economic models lack interactions between agents, such as trading, and consider simple, intertemporal dynamics. However, these theoretical approaches do not yield a complete tax policy in more complex settings, such as the Gather-Trade-Build environment studied in this work. For instance, one can derive analytical expressions for the distortion in savings or capital due to taxation ( 36), or the asymptotic behavior of taxes in the far future, the tax rate on the highest-skill agent, or the structure of incentives ( 37). In stylized multistep settings, analytical methods have studied implicitly characterizations of the impact of taxes. Concrete results are only available for two-step economies ( 35). Although work in New Dynamic Public Finance ( 33, 34) seeks to model multistep economies, these models quickly become challenging to study analytically. ![]() For example, typical models use static, one-step economies ( 31, 32) and make use of assumptions about people’s sensitivity to tax changes (elasticity). Theory-driven approaches to tax policy design have needed to make simplifications in the interest of analytical tractability ( 30). In contrast, using deep RL is substantially less constrained, i.e., the neural networks can implement a very high-dimensional space of policies and thus model a large and diverse set of policies. Moreover, it is difficult to compute equilibria across any given set of policies, even if EGTA-style algorithms can sample from (high-dimensional) policy spaces. ![]() However, it is hard to apply such methods to our spatiotemporal economies, which are substantially more complex than traditionally studied games and use an unbounded space of policies. Machine learning methods have been used to enhance EGTA, e.g., policy space response oracle methods where an oracle returns the best response(s) to (a set of meta-game) agent policies ( 25, 26). EGTA generally constrains agent policies to a predefined set, with replicator dynamics used to find equilibria across these heuristic policies (for symmetric games with interchangeable agents and identical affordances) ( 24). A computational approach to finding equilibria is empirical game-theoretic analysis (EGTA). However, understanding the landscape of equilibria in general-sum games remains a substantial challenge ( 4). These results demonstrate that two-level, deep RL complements economic theory and unlocks an AI-based approach to designing and understanding economic policy.Įconomies with multiple agents and a social planner can also be seen as a hierarchical general-sum game, in which the planner designs payoffs for which the agents optimize their behavior. It does so despite emergent tax-gaming strategies while accounting for emergent labor specialization, agent interactions, and behavioral change. In spatiotemporal economies, the AI Economist substantially improves both utilitarian social welfare and the trade-off between equality and productivity over baselines. In one-step economies, the AI Economist recovers the optimal tax policy of economic theory. We validate this framework in the domain of taxation. In particular, the AI Economist uses structured curriculum learning to stabilize the challenging two-level, coadaptive learning problem. The AI Economist is a two-level, deep RL framework for policy design in which agents and a social planner coadapt. Artificial intelligence (AI) and reinforcement learning (RL) have improved many areas but are not yet widely adopted in economic policy design, mechanism design, or economics at large. ![]()
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |