Saved in:
| Main Authors: | , , , |
|---|---|
| Format: | Preprint |
| Published: |
2024
|
| Subjects: | |
| Online Access: | https://arxiv.org/abs/2402.08747 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| _version_ | 1866913233803149312 |
|---|---|
| author | Bajaj, Shivam Das, Pranoy Vorobeychik, Yevgeniy Gupta, Vijay |
| author_facet | Bajaj, Shivam Das, Pranoy Vorobeychik, Yevgeniy Gupta, Vijay |
| contents | Many learning algorithms are known to converge to an equilibrium for specific classes of games if the same learning algorithm is adopted by all agents. However, when the agents are self-interested, a natural question is whether agents have a strong incentive to adopt an alternative learning algorithm that yields them greater individual utility. We capture such incentives as an algorithm's rationality ratio, which is the ratio of the highest payoff an agent can obtain by deviating from a learning algorithm to its payoff from following it. We define a learning algorithm to be $c$-rational if its rationality ratio is at most $c$ irrespective of the game. We first establish that popular learning algorithms such as fictitious play and regret matching are not $c$-rational for any constant $c\geq 1$. We then propose and analyze two algorithms that are provably $1$-rational under mild assumptions, and have the same properties as (a generalized version of) fictitious play and regret matching, respectively, if all agents follow them. Finally, we show that if an assumption of perfect monitoring is not satisfied, there are games for which $c$-rational algorithms do not exist, and illustrate our results with numerical case studies. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2402_08747 |
| institution | arXiv |
| publishDate | 2024 |
| record_format | arxiv |
| spellingShingle | Rationality of Learning Algorithms in Repeated Normal-Form Games Bajaj, Shivam Das, Pranoy Vorobeychik, Yevgeniy Gupta, Vijay Computer Science and Game Theory Systems and Control Many learning algorithms are known to converge to an equilibrium for specific classes of games if the same learning algorithm is adopted by all agents. However, when the agents are self-interested, a natural question is whether agents have a strong incentive to adopt an alternative learning algorithm that yields them greater individual utility. We capture such incentives as an algorithm's rationality ratio, which is the ratio of the highest payoff an agent can obtain by deviating from a learning algorithm to its payoff from following it. We define a learning algorithm to be $c$-rational if its rationality ratio is at most $c$ irrespective of the game. We first establish that popular learning algorithms such as fictitious play and regret matching are not $c$-rational for any constant $c\geq 1$. We then propose and analyze two algorithms that are provably $1$-rational under mild assumptions, and have the same properties as (a generalized version of) fictitious play and regret matching, respectively, if all agents follow them. Finally, we show that if an assumption of perfect monitoring is not satisfied, there are games for which $c$-rational algorithms do not exist, and illustrate our results with numerical case studies. |
| title | Rationality of Learning Algorithms in Repeated Normal-Form Games |
| topic | Computer Science and Game Theory Systems and Control |
| url | https://arxiv.org/abs/2402.08747 |