Dynamic Restart Policies

1
Dynamic Restart Policies

• JIN Xiaolong

(Based on 1 )
2
Traditional Restart

• Two assumptions
• Only one feasible observation the length of a
  run
• Complete or no knowledge
• Results
• For complete knowledge a fixed cutoff
• For no knowledge 1, 1, 2, 1, 1, 2, 4,

3
Dynamic Restart

• Basic Idea
• Sample from an instance ensemble
• Explore the possible run-time distribution(s) of
  the samples by a Bayesian Learning Method
• Build a predictive model to predict run-time

4
Dynamic Restart (Cont.)
One possible distribution
Multiple possible distributions
T the cutoff q(t) the probability of a run
stopping at t E(T) the expected run-time given
a cutoff T
5
Results Features

• Dynamic restart policies VS Fixed restart
  policies
• about 4065 improvement in solution time
• Features
• Dynamically tune the cutoff
• The running is separated with the learning
• Bayesian learning method is slow in speed

6
References
1. H. Kautz, E. Horvitz, Y. Ruan, C. Gomes, B.
Selman. Dynamic Restart Policies, Proceedings of
the Eighteenth National Conference on Artificial
Intelligence, Edmonton, Alberta, July 2002. AAAI
Press.