Exit time analysis of Kesten's stochastic recurrence equation and SGD, analysis on the convergence rate of CFR+ algorithm
김선우
27동 220호
0
104
06.02 11:02
| 구분 | 박사학위 논문 발표 |
|---|---|
| 일정 | 2025-06-11(수) 10:00~11:00 |
| 세미나실 | 27동 220호 |
| 강연자 | 유지호 (서울대학교) |
| 담당교수 | 서인석 |
| 기타 |
발표자 유지호
일시: 2025.06.11. 10:00 ~ 10:50
제목: Exit time analysis of Kesten's stochastic recurrence equation and SGD, analysis on the convergence rate of CFR+ algorithm
초록: In this lecture, we review classic theory on the Kesten's stochastic recurrence equation which naturally appears in the analysis of various time series such as SGD or GARCH. We then explain our recent result on the escape time analysis on two major regimes which connects the Lyapunov exponent of the process and the escape time from a domain. Then, we try to connect our result to the empirically observed behavior of the time series quoted above. Plus, we introduce optimization algorithms called CFR and its variant CFR+, widely used to find Nash equilibriums in incomplete information games such as Heads up limit Texas Hold'em(HULHE). We determined the exact time scale of regret convergence of CFR+ in a two-player toy model - an open question since the algorithm's introduction.
