220 likes | 319 Vues
This experiment introduces the STAGE algorithm by Boyan for optimizing objective functions, improving upon WalkSAT results. Utilizing value function approximation and feature Markovianization, STAGE enhances problem-solving methods. Parameters such as noise, patience, and cutoff values are crucial for effective tuning. The function approximator V-bar-pi, using linear regression, produces more efficient results. Discussions on feature coefficients, time evaluations, learning concepts, and random restarts provide insights into STAGE's performance compared to WalkSAT. The experiment also explores solving Hanoi problems with carefully tuned parameters.
E N D
Experiments with STAGE Wei Wei
Introduction • STAGE- Developed by Boyan • Use value function approximation to automatically analyze sample trajectories. • Speed up many local search methods
Diagram of STAGE Produces new training data Run p to Optimize Obj Hillclimb to Optimize V Produces good start states
Apply it to SAT • The base algorithm is WalkSAT (modified) • Got results better than pure WalkSAT
Overview • We need to deal with four aspects of the problem: WalkSAT, STAGE, features, and to make the algorithm Markovian. • Hard to tune; not every combination works. stage WalkSAT features Marko-vianize
Features • %clauses unsatisfied (-) • %clauses satisfied by 1 variable (+) • %clauses satisfied by 2 variables (-) • %critical variables (-) • %variables set to naïve setting (~)
Markovianize • S/W1 : patience based, not Markovian • S/W2 : best-so-far • S/W3 : epsilon cutoff
Parameter tuning • Noise 0.25 seems good • Patience 10,000 • Cutoff 1,000,000 • Epsilon .0001
Function approximator V-bar-pi • Quadratic regression • Linear regression • Linear functions perform 25% better, and faster. • Linear functions are coarse approximators.
Discussion(1) • Linear regression is very bad approximation is this case, yet it gives better results than quadratic regression. Why? • Hit bottom very often • Lead to long more WalkSAT moves
Discussion(2) • Features – coefficients vary a lot among instances. But relatively stable within one instance. • The signs are relatively stable
Discussion(3) • Time vs evaluation • When # of evaluation is fixed, STAGE performs 3 times better, but time spent is doubled • When time is fixed, the result is 40% better than WalkSAT
Discussion(4) • Can it hit the finish line? • It does vaguely(?) learn some concepts, which hopefully can direct WalkSAT to a good place. • Par-? Is a good set of problems to solve?
One feature No improvement over WalkSAT.
Random restart • 176 Random flips – Worse than S/W3, still better than WalkSAT • 1000 Random flips – Worse than one-run WalkSAT • Complete new start points – similar to the case above. • Parameters: cutoff – 10,000. Restart – 100.
Hanoi • Parameters not yet carefully tuned • It would be interesting to see whether Hanoi4 can be solved by carefully tuned S/W3. I ran WalkSAT for 50,000,000 flips, but failed to solve it.