论序贯激励设计的复杂性
On the Complexity of Sequential Incentive Design
摘要:
在许多情况下,委托人动态地与代理人交互,并提供一系列激励措施,以使代理人的行为与期望的目标相一致。本文重点研究的问题,合成的激励序列,一旦提供,诱导所需的代理人的行为,即使代理人的内在动机是未知的主要。我们将代理的行为建模为马尔可夫决策过程,将其内在动机表示为奖励函数,该奖励函数属于有限的可能报酬函数,并将激励视为提供给代理的额外奖励。我们首先表明,行为修改问题(BMP),即,综合激励序列的问题,诱导所需的代理人的行为,在最小的总成本的委托人,是PSPACE困难的。此外,我们表明,通过对主要的激励序列施加一定的限制,可以得到两个NP-完全的BMP变体。我们还提供了一个充分条件的一组可能的报酬函数下,BMP可以解决通过线性规划。最后,我们提出了两种算法来计算BMP的NP完全变体的全局和局部最优解。
英文摘要:
In many scenarios, a principal dynamically interacts with an agent and offers a sequence of incentives to align the agent's behavior with a desired objective. This paper focuses on the problem of synthesizing an incentive sequence that, once offered, induces the desired agent behavior even when the agent's intrinsic motivation is unknown to the principal. We model the agent's behavior as a Markov decision process, express its intrinsic motivation as a reward function, which belongs to a finite set of possible reward functions, and consider the incentives as additional rewards offered to the agent. We first show that the behavior modification problem (BMP), i.e., the problem of synthesizing an incentive sequence that induces a desired agent behavior at minimum total cost to the principal, is PSPACE-hard. Moreover, we show that by imposing certain restrictions on the incentive sequences available to the principal, one can obtain two NP-complete variants of the BMP. We also provide a sufficient condition on the set of possible reward functions under which the BMP can be solved via linear programming. Finally, we propose two algorithms to compute globally and locally optimal solutions to the NP-complete variants of the BMP.
机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。