Human-Guided Hill Climbing for Gallery Wall Design

Class Discussion Prompt

Today we are going to think about how a human curator might guide an algorithm to iteratively improve the selection and arrangement of artworks on a gallery wall.

So far, we have been thinking about algorithms that generate candidate walls and score them automatically. But in a real gallery setting, there often is no single fixed definition of “best.” What counts as a good wall depends on the goals of the show, the taste of the curator, and the effect the gallery wants to create.

This leads us to a different model:

Instead of treating the algorithm as something that finds the one objectively best answer, we can treat it as a system that proposes possibilities, receives human feedback, and then tries to improve its next proposal.

In other words, the human is not outside the search process. The human becomes part of the search process.

The Basic Idea

Suppose we have a scoring.yaml file that describes what this particular gallery show cares about.

Rather than hard-coding one permanent notion of quality, the scoring file might contain a set of parameters such as:

  • thematic unity vs thematic diversity
  • color harmony vs contrast
  • historical consistency vs mixing of periods
  • emotional consistency vs emotional tension
  • balance of sizes
  • smoothness of transitions from one artwork to the next
  • how strongly a wall should center around a focal piece

Each of these could have:

  • a description
  • a range
  • a current preferred value
  • an importance weight

This would mean that the system is not searching for the best wall in an absolute sense. It is searching for the best wall given the current preferences of the curator.

Example

A curator might say:

  • “For this show, I want the works to feel strongly connected by theme.”
  • Or instead: “I want variety and tension. I do not want everything to feel too similar.”

Those are very different goals.

So a parameter in the scoring file might not simply say “more theme is better.” Instead, it might represent a scale:

  • one end = strong thematic contrast
  • the other end = strong thematic unity

Then the curator can say where on that scale they want this show to fall.

This is important because many criteria are not just “maximize this.” They are really “aim for this region of the scale.”

Why This Matters

A fixed evaluator assumes we already know exactly what makes a wall good.

But in practice:

  • one show may want unity
  • another may want contrast
  • one may want calm transitions
  • another may want tension and surprise
  • one may want works from the same period
  • another may want periods intentionally mixed

So instead of building a single hard-coded evaluator, we are really building a configurable evaluator.

That makes the system much more realistic and much more interesting.

Human-Guided Hill Climbing

Ordinary hill climbing works like this:

  1. Generate a candidate solution
  2. Score it
  3. Make a small change
  4. Keep the new version if it scores better

In this project, the process might become:

  1. Generate a wall
  2. Show it to the curator
  3. The curator reacts to it
  4. Update the scoring preferences
  5. Generate a revised wall
  6. Continue iterating

So the system is not just improving the wall. It is also learning what kind of wall the human wants.

This means the objective function is no longer fixed. It can evolve as the human sees examples and refines their thinking.

That is a much richer problem than ordinary hill climbing.

A Useful Way to Think About the Problem

There are really three layers here.

1. Artwork Features

What properties does each artwork have?

For example:

  • theme
  • period
  • medium
  • size
  • color palette
  • emotional tone
  • style

2. Wall Metrics

Given a complete wall, what do we measure?

For example:

  • average thematic similarity
  • amount of contrast between neighboring works
  • size balance
  • palette harmony
  • chronological spread
  • emotional pacing across the wall

3. Curator Preferences

What does this particular show want?

For example:

  • preferred level of thematic unity
  • preferred amount of contrast
  • importance of color harmony
  • importance of historical range
  • acceptable tolerance around preferred values

This separation is useful because it helps us think clearly about what the system knows, what it measures, and what the human is controlling.

Discussion Questions

Please come prepared to discuss the following questions.

1. What should go into the scoring file?

What are the most important parameters a gallery might care about for a show?

Think broadly. These could include visual factors, thematic factors, historical factors, emotional factors, or something else.

2. Which parameters are naturally opposites on a scale?

For example:

  • unity vs diversity
  • harmony vs contrast
  • consistency vs surprise

Are there parameters that should be treated as a continuum rather than as something to simply maximize?

3. What kind of human input should the curator provide after each iteration?

Should the curator:

  • adjust sliders for preferences?
  • rate the current wall on several dimensions?
  • check boxes saying what improved or got worse?
  • lock in certain artworks they want to keep?
  • identify pairs that should or should not be adjacent?
  • choose between two candidate walls?

What kind of feedback would be both useful and realistic?

4. What counts as a “neighbor” in hill climbing for this problem?

What small changes should the system be allowed to try?

For example:

  • swap out one artwork
  • reorder two artworks
  • move one artwork to a new position
  • replace a cluster of works
  • keep a favorite work fixed and optimize around it

Which kinds of changes seem most useful?

5. Which criteria depend on selection, and which depend on arrangement?

Some qualities depend mostly on which artworks are chosen.

Others depend mostly on where they are placed relative to one another.

Can you identify examples of each?

6. Should the evaluator reward hitting a target rather than just maximizing a score?

Suppose the curator wants “moderate contrast” rather than “maximum contrast.”

How should that be represented?

Should the system try to get close to a preferred value rather than simply make a metric as large as possible?

7. How should the human guide the search?

Should the human be mainly adjusting the scoring model, or directly judging candidate walls, or both?

What would make the process understandable and effective?

Questions to Push Further

If time permits, also think about these:

  • Could the system become too responsive to short-term human reactions?
  • How should we prevent the search from getting stuck in a narrow local pattern?
  • How much control should the human have over the scoring model?
  • Should the system remember previous curator responses and try to infer broader preferences?
  • At what point does this become less like pure optimization and more like interactive design?

Goal for Today’s Discussion

By the end of today’s discussion, we want to begin forming a clearer idea of:

  1. what the scoring.yaml file should contain,
  2. what kinds of human feedback the interface should support,
  3. what kinds of moves the hill-climbing algorithm should make,
  4. and how the human and the algorithm should work together to improve a wall over time.

The goal is not just to make an algorithm that scores walls.

The goal is to think about how to build an algorithmic system that supports human judgment, changing preferences, and iterative improvement.

Closing Thought

A good gallery wall is not defined by a universal formula.

It is defined by a set of curatorial goals, some of which may change as the curator sees new possibilities.

Our challenge is to design a system that can represent those goals, generate candidate walls, accept human guidance, and iteratively improve the result.

Scroll to Top