Kappa decay #221
Kappa decay #221
Conversation
Add a tunable parameter kappa decay that will allow you to make the optimization more exploitative over time. From a few tests I've run i get faster convergence to better maxima.
|
Thanks for the PR. I really like the idea of adding this. However I'm not a fan of recreating the |
bayes_opt/bayesian_optimization.py
Outdated
| @@ -162,9 +163,16 @@ def maximize(self, | |||
| self._prime_queue(init_points) | |||
| self.set_gp_params(**gp_params) | |||
|
|
|||
| util = UtilityFunction(kind=acq, kappa=kappa, xi=xi) | |||
| def decay_kappa(kappa, iters, decay): | |||
There was a problem hiding this comment.
Can we turn this into a UtilityFunction method?
bayes_opt/bayesian_optimization.py
Outdated
| iteration = 0 | ||
| while not self._queue.empty or iteration < n_iter: | ||
| bo_iters = max(0, len(self.res) - init_points) | ||
| decayed_kappa = decay_kappa(init_kappa, bo_iters, kappa_decay) | ||
| util = UtilityFunction(kind=acq, kappa=decayed_kappa, xi=xi) |
There was a problem hiding this comment.
Use the method to update the value here as opposed to creating a new object.
There was a problem hiding this comment.
I put the decay in UtilityFunction and also added an optional "delay" parameter to give users the option to delay kappa decay by X BO iters before kappa starts to decay.
I think this will be useful for ensuring a fixed amount of explorative BO in the beginning steps before decaying kappa to be more exploitative. As with original kappa decay it is optional and by default turned off so most users will not even know this is an option. Doesn't change the existing API either.
Lmk what you think.
also add an additional parameter if the user wants to delay the kappa decay for a certain amount of BO iterations
Codecov Report
@@ Coverage Diff @@
## master #221 +/- ##
==========================================
- Coverage 99.68% 98.77% -0.92%
==========================================
Files 6 6
Lines 322 327 +5
Branches 32 33 +1
==========================================
+ Hits 321 323 +2
- Misses 0 2 +2
- Partials 1 2 +1
Continue to review full report at Codecov.
|
bayes_opt/util.py
Outdated
| @@ -99,6 +99,9 @@ def utility(self, x, gp, y_max): | |||
| return self._ei(x, gp, y_max, self.xi) | |||
| if self.kind == 'poi': | |||
| return self._poi(x, gp, y_max, self.xi) | |||
|
|
|||
| def decay_kappa(self, decay_factor): | |||
There was a problem hiding this comment.
I meant passing the decay_factor (and _delay too) to the UtilityFunction constructor, and then calling util.update_params() (or similar name) in the loop. All this logic if kappa_decay != 1 and max(0, len(self.res) - init_points) - kappa_decay_delay > 0: could reside inside the UtilityFunction object.
There was a problem hiding this comment.
Ah i see. I will make the changes. For delay to be used however we will need to feed in the # of BO iters to the update function. Lmk how it looks now.
There was a problem hiding this comment.
UtilityFunction could keep an internal counter to keep track of delay which is updated every time the method for kappa decay is invoked.
bayes_opt/bayesian_optimization.py
Outdated
| @@ -165,6 +167,10 @@ def maximize(self, | |||
| util = UtilityFunction(kind=acq, kappa=kappa, xi=xi) | |||
| iteration = 0 | |||
| while not self._queue.empty or iteration < n_iter: | |||
| # decay kappa by decay factor if BO iters exceeds kappa_decay_delay | |||
| if kappa_decay != 1 and max(0, len(self.res) - init_points) - kappa_decay_delay > 0: | |||
There was a problem hiding this comment.
I rather not introduce this if statement here. It's at a different abstraction level than the surrounding code and it also only applies to one kind of utility function.
There was a problem hiding this comment.
is this still something you wanted to move forward with?
There was a problem hiding this comment.
Yes, absolutely. I've been a bit busy lately with work, so I'm slow to make comments and stuff. But we'll get this merged this week =)
|
The implementation is great! I'm ready to approve, but I have one small ask first. Would you mind rebasing and squashing all commits in a single one? |
Add a tunable parameter kappa decay that will allow you to make the optimization more exploitative over time. From a few tests I've run i get faster convergence to better maxima. move decay to a UtilityFunction method also add an additional parameter if the user wants to delay the kappa decay for a certain amount of BO iterations add decay to UtilityFunction Update bayesian_optimization.py Update bayesian_optimization.py Update bayesian_optimization.py Update bayesian_optimization.py Update util.py Update util.py Update util.py Update util.py Update bayesian_optimization.py Update util.py Update bayesian_optimization.py Update bayesian_optimization.py
parent 19a2314 author ben-arnao <arnaoben@gmail.com> 1588781383 -0400 committer ben-arnao <arnaoben@gmail.com> 1588788584 -0400 Kappa decay Add a tunable parameter kappa decay that will allow you to make the optimization more exploitative over time. From a few tests I've run i get faster convergence to better maxima. move decay to a UtilityFunction method also add an additional parameter if the user wants to delay the kappa decay for a certain amount of BO iterations add decay to UtilityFunction Update bayesian_optimization.py Update bayesian_optimization.py
|
Hey guys, this is a really cool feature, but it took me a while to realise it was here - unless I missed it I don't think this is covered in the examples? |
Add a tunable parameter kappa decay that will allow you to make the optimization more exploitative over time. From a few tests I've run i get faster convergence to better maxima.