Pareto optimization #475

AdrianSosic · 2025-02-03T12:35:23Z

This PR finally brings Pareto optimization via the new ParetoObjective class, together with an example comparing Pareto vs. single target optimization for a simple pair of synthetic targets.

Note: Support is currently limited to maximization and minimization targets. Match targets will follow but require a refactoring of the corresponding target transformation mechanism.

README.md

examples/Multi_Target/pareto.py

Scienfitz

first round of comments

CHANGELOG.md

baybe/acquisition/acqfs.py

baybe/acquisition/__init__.py

README.md

examples/Multi_Target/pareto.py

CHANGELOG.md

docs/userguide/objectives.md

baybe/surrogates/composite.py

Co-authored-by: Martin Fitzner <[email protected]>

The ref_point is now in the original target space so that the user can intuitively specify its coordinates. Sign flips for minimization targets happen behind the scenes.

Scienfitz · 2025-02-17T21:16:42Z

baybe/recommenders/pure/bayesian/base.py

+        ):
+            raise IncompatibleAcquisitionFunctionError(
+                f"You attempted to use a single-target acquisition function in a "
+                f"{n_targets}-target context."


the language here is not good due to the same issue elsewhere: a single target acqf is perfectly fine eg for a desirability objective (ie multiple targets), whats really the problem is the number of outputs (and not targets)

Scienfitz · 2025-02-17T21:25:02Z

CHANGELOG.md

detached comment 3: working with Pareto I think the Campaign.posterior call is much more important and probaly part fo the workflow (to check the target predictions and possibly make a subchoice of poitns to evaluate on the predicted frontier). Imo this then should be mentioned both in the UG and in the example (albeit briefy, jsut referencing the posterior method and explaiing why its useful for that)

baybe/surrogates/composite.py

Scienfitz · 2025-02-17T21:31:18Z

baybe/surrogates/base.py

+        # Validate multi-target compatibility
+        if not self.supports_multi_target and (n_targets := len(objective.targets)) > 1:
+            raise IncompatibleSurrogateError(
+                f"You attempted to train a single-target surrogate in a "


Suggested change

f"You attempted to train a single-target surrogate in a "

f"You attempted to train a single-output surrogate in a "

Scienfitz · 2025-02-17T21:32:49Z

baybe/surrogates/composite.py

+
+
+@define
+class BroadcastingSurrogate(SurrogateProtocol):


I was expecting a supports_multi_target=True in this class but perhaps its not needed or you had other reasons?

Yesterday I answered "yes, thanks, forgot" but I just realize that I omitted it on purpose since actually not needed / fulfills no purpose here. The potential confusion is: note that the two composite classes are not subclasses of Surrogate (which is where the flag is declared), they only implement the Protocol. Because the flag is only used inside Surrogate, there is no benefit of also exposing a copy of it in the composite ones – they will at some point dispatch the the actual surrogates, which is where the check is then performed (and if coming via this route, it will not even be required since the number of encountered targets is then trivially 1). So it's only ever needed if you don't go via the composite surrogates. Makes sense? If yes, please resolve

I just thought having such flag available outside of Surrogate could be useful, think eg about tests

Scienfitz · 2025-02-18T10:51:48Z

baybe/acquisition/acqfs.py

+        The reference point is positioned in relation to the worst target configuration
+        within the provided array. The distance in each target dimension is adjusted by
+        a specified multiplication factor, which scales the reference point away from
+        the worst target configuration based on the maximum observed differences in
+        target values.


Suggested change

The reference point is positioned in relation to the worst target configuration

within the provided array. The distance in each target dimension is adjusted by

a specified multiplication factor, which scales the reference point away from

the worst target configuration based on the maximum observed differences in

target values.

The reference point is positioned relative to the worst point in the direction coming from the best

point. A factor of 0.0 would result in the reference point being the worst point, while a factor > 0.0

would move the reference point further away from both worst and best points. A factor of 1.0 would

exactly mirror the best on on the worst point.

Scienfitz · 2025-02-18T10:59:21Z

examples/Multi_Target/pareto.py

+plt.ylabel(y1.name)
+plt.title("Target Space")
+plt.axis("square")
+


can we mix target types in this example? ie one max and one min
reason: its safer to test mixed targets as both cases for the multiplier are contained

Scienfitz · 2025-02-18T11:01:16Z

docs/userguide/objectives.md

+
+```{admonition} Non-dominated Configurations
+:class: tip
+A target configuration is considered non-dominated if no other configuration is better 


for a newbie otherwise its probably not clear why the objective is even called Pareto

Suggested change

A target configuration is considered non-dominated if no other configuration is better

A target configuration is considered non-dominated (or Pareto-optimal) if no other configuration is better

Scienfitz · 2025-02-18T11:06:35Z

baybe/surrogates/composite.py

+
+
+@define
+class CompositeSurrogate(SurrogateProtocol):


is there any merit to deriving BroadcastingSurrogate from CompositeSurrogate?
It seems like a special case to me

No, inheritance does not help here – the attribute layout and way of calling the initializer is different. But I know what you had in mind: a BroadcastingSurrogate indeed behaves like a CompositeSurrogate with a flexible collection of single-task surrogates. So what you probably meant – and what indeed work – is to compose the CompositeSurrogate in that case via an alternative construction route:

@classmethod def from_template(cls, surrogate: SurrogateProtocol) -> CompositeSurrogate: return cls(defaultdict(lambda: deepcopy(surrogate)))

With that, doing a BroadcastingSurrogate(template) would be identical to CompositeSurrogate.from_template(template).

In fact:

the amount of code required would be significantly less (i.e. one entire class dropped)

everything should still work via the API since we have our clever constructor flag

we could still use the auto-conversion for single-target surrogates

But the consequence is that the creation step would no longer be a regular "class call" but requires this separate route instead. So which do we prefer?

Scienfitz · 2025-02-18T11:09:53Z

baybe/surrogates/base.py

+        if not self.supports_multi_output and (n_targets := len(objective.targets)) > 1:
+            raise IncompatibleSurrogateError(
+                f"You attempted to train a single-target surrogate in a "
+                f"{n_targets}-target context. Either use a proper multi-target "


Suggested change

f"{n_targets}-target context. Either use a proper multi-target "

f"context requiring {n_targets} model outputs. Either use a proper multi-output "

Scienfitz · 2025-02-18T11:30:12Z

baybe/surrogates/composite.py

+
+
+@define
+class BroadcastingSurrogate(SurrogateProtocol):


not mentioned anywhere in the surrogate userguide so far, but I think these are important explanations.

Im not calling for a complete revamp, but at least mention that the models currently mentioned are all single-output (with the exception of GP where it uses the builtin botorch mechanism)

and then mention the current ways of creating multi-output models via composite or broadcastin from the above

long term (not for this PR) I can imagine a section explaining (perhaps even with some flow charts) the four multi-output possibilities

botorch builtin broadcasting (comp efficient, no partial measurements)

broadcasting a single-target model, supports partial measurements

composite, supports partial targets

an inherently multi-output model not consisting of any independent models

do I understand it right that after this PR we already have possibilities 1 to 3 (without the partial targets of course)

Scienfitz · 2025-02-18T11:37:59Z

baybe/objectives/__init__.py

besides acqfs ad surrogates, what about also giving all objectives a property supports_multi_output ?

if we have a reason for it –> why not. but if not, why add unnecessary code to maintain? what use case do you have in mind?

use case in mind is mostly tests and any kind of sutiation where i need to distinguish that property programatically
i also think the objectives are prob one of the easiest to maintain this flag

I could imagine for instance doc pages that automatically iterate over avaialble objects,g rouping them into multi-output / single-output etc
the other usecase is always tests

OK 👍🏼 but in that case I suggest we do the usual thing: we add it in that second when it's needed, instead of doing the potentially this might be helpful-logic

AdrianSosic added the new feature New functionality label Feb 3, 2025

AdrianSosic self-assigned this Feb 3, 2025

AdrianSosic requested review from Scienfitz and AVHopp as code owners February 3, 2025 12:35

AdrianSosic commented Feb 3, 2025

View reviewed changes

README.md Show resolved Hide resolved

examples/Multi_Target/pareto.py Outdated Show resolved Hide resolved

AdrianSosic force-pushed the feature/pareto branch 2 times, most recently from 385bebe to e633f92 Compare February 3, 2025 14:08

Scienfitz reviewed Feb 11, 2025

View reviewed changes

AdrianSosic force-pushed the feature/pareto branch 5 times, most recently from 27ae08c to 711935f Compare February 13, 2025 15:22

Scienfitz reviewed Feb 13, 2025

View reviewed changes

docs/userguide/objectives.md Outdated Show resolved Hide resolved

docs/userguide/objectives.md Show resolved Hide resolved

AdrianSosic commented Feb 17, 2025

View reviewed changes

baybe/surrogates/composite.py Show resolved Hide resolved

AdrianSosic force-pushed the feature/pareto branch from 44956f9 to 01dbb6f Compare February 17, 2025 08:07

AdrianSosic added 14 commits February 17, 2025 13:39

Improve deprecation warning message

bdd1e90

Draft ParetoObjective class

5e52039

Extract function for transforming target columns

460bbc1

Add qLogNEHVI acqusition function

04056ee

Make botorch multiobjective acqusition functions autodetectable

8b2d199

Add temporary restriction allowing only MAX targets

6788abb

Draft example

1043156

Enable minimization targets

473ac15

Add highlighted feature to README

41542f1

Update CHANGELOG.md

d0d7716

Compute default reference point from data

75be534

Flip signs of custom reference points in MIN mode

0e340e5

Interpolate target paretor frontier along transformed points

12eef75

Drop unnecessary label arguments

0f29bda

AdrianSosic and others added 24 commits February 17, 2025 13:39

Dynamically select default acquisition function

7b97131

Deactivate comparison for non-persistent attributes

eb6aad6

Fix variable reference in example

9cf8363

Co-authored-by: Martin Fitzner <[email protected]>

Turn assert statement into proper exception

5795a4a

Add prune_baseline attribute

f4384bd

Add full docstring to compute_ref_point

1fd25e9

Refactor ref_point computation logic

0c605b8

The ref_point is now in the original target space so that the user can intuitively specify its coordinates. Sign flips for minimization targets happen behind the scenes.

Let doc generation append regular image when available

81bf88c

Add ParetoObjective user guide section

f3e5c57

Add surrogate broadcasting mechanism

3b41310

Validate multi-target compatibility

3deee1b

Rename broadcasting.py to composite.py

0252ef0

Add CompositeSurrogate class

26c7e36

Add surrogate composition test

d22bdac

Add TODO note

3e867d2

Add qLogNoisyExpectedHypervolumeImprovement to strategy

966cfd7

Add missing strategy arguments

43872a2

Add pareto_objectives strategy and serialization test

c157d01

Fix default acquisition function mechanism

3383fd9

Throw exception when using single-target acqf in multi-target context

5918719

Use specific incompatibility errors instead of generic ValueError

727ac31

Drop opinionated statement from user guide

d124ffd

Mention requirement of multi-target acquisition function in user guide

78fee0c

Update CHANGELOG.md

b3b77fb

AdrianSosic force-pushed the feature/pareto branch from 01dbb6f to b3b77fb Compare February 17, 2025 12:39

AdrianSosic added 3 commits February 18, 2025 10:03

Rename supports_multi_target to support_multi_output

dc3d601

Fix Liskov

f8b6b61

Ignore typing problem in classproperty

ec9570f

Scienfitz reviewed Feb 18, 2025

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Pareto optimization #475

Pareto optimization #475

AdrianSosic commented Feb 3, 2025

Scienfitz left a comment

Scienfitz Feb 17, 2025

Scienfitz Feb 17, 2025

Scienfitz Feb 17, 2025

Scienfitz Feb 17, 2025

AdrianSosic Feb 19, 2025

Scienfitz Feb 19, 2025

Scienfitz Feb 18, 2025

Scienfitz Feb 18, 2025

Scienfitz Feb 18, 2025

Scienfitz Feb 18, 2025

AdrianSosic Feb 19, 2025

Scienfitz Feb 18, 2025

Scienfitz Feb 18, 2025

Scienfitz Feb 18, 2025

AdrianSosic Feb 19, 2025

Scienfitz Feb 19, 2025

Scienfitz Feb 19, 2025

AdrianSosic Feb 19, 2025

	f"You attempted to train a single-target surrogate in a "
	f"You attempted to train a single-output surrogate in a "

	A target configuration is considered non-dominated if no other configuration is better
	A target configuration is considered non-dominated (or Pareto-optimal) if no other configuration is better

	f"{n_targets}-target context. Either use a proper multi-target "
	f"context requiring {n_targets} model outputs. Either use a proper multi-output "

Pareto optimization #475

Are you sure you want to change the base?

Pareto optimization #475

Conversation

AdrianSosic commented Feb 3, 2025

Scienfitz left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment