Feature: Simulated annealing optimization by VoX · Pull Request #1 · VoX/Bob-Rust-Java

VoX · 2026-04-04T21:06:49Z

Summary

Replace pure hill-climbing shape refinement with simulated annealing (SA) that escapes local minima by probabilistically accepting worse moves early in the search
Keep the old hill climbing accessible via USE_SIMULATED_ANNEALING flag in AppConstants (defaults to true)
Increase parallel SA chains (times in Model.java) to availableProcessors / 2 for better exploration

Changes

File	Change
`HillClimbGenerator.java`	Add `getHillClimbSA()`, `estimateTemperature()`, `computeCoolingRate()`; rename old method to `getHillClimbClassic()`; dispatch via feature flag
`Model.java`	`times` now uses `Runtime.getRuntime().availableProcessors() / 2` instead of hardcoded `1`
`AppConstants.java`	Add `USE_SIMULATED_ANNEALING` boolean flag
`SimulatedAnnealingBenchmark.java`	JUnit 5 tests: temperature schedule validation, SA vs HC energy comparison, regression checks on multiple image types
`TestImageGenerator.java`	Programmatic 128x128 test image generation (solid, gradient, edges, photo detail, nature)

How SA works

Temperature estimation: Sample 30 random mutations, measure average energy delta, set initial temperature so ~60% of uphill moves are accepted
Cooling schedule: Geometric decay computed so temperature reaches 0.001 after maxAge * 10 iterations
Accept/reject: Improvements always accepted; worse moves accepted with probability exp(-delta / T); best-seen state tracked and returned

Test plan

testTemperatureSchedule — verifies temperature is positive/finite, cooling rate is in (0,1), final temp near zero
testSAProducesLowerOrEqualEnergy — SA score within 5% of HC on photo detail image
testSANeverSignificantlyWorse — SA within 5% of HC on solid, gradient, and edges images
testCoolingRateEdgeCases — handles tiny and large initial temperatures
./gradlew clean build passes cleanly

🤖 Generated with Claude Code

Replace the pure hill-climbing refinement in HillClimbGenerator with simulated annealing (SA) that can escape local minima by probabilistically accepting worse moves early in the search. The old hill climbing is kept as a fallback behind the USE_SIMULATED_ANNEALING flag in AppConstants. Key changes: - HillClimbGenerator: add getHillClimbSA(), estimateTemperature(), and computeCoolingRate() methods; dispatch via feature flag in getHillClimb() - Model: increase parallel SA chains (times) to use availableProcessors/2 - AppConstants: add USE_SIMULATED_ANNEALING boolean flag (default true) - Add JUnit 5 benchmark tests comparing SA vs classic hill climbing - Add programmatically generated test images (128x128) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Copilot

Pull request overview

This PR introduces a simulated annealing (SA) based optimizer for shape refinement to better escape local minima, while keeping the classic hill climb path available behind a feature flag.

Changes:

Added SA optimizer (getHillClimbSA) with temperature estimation and geometric cooling; classic hill climb retained as getHillClimbClassic and dispatched via USE_SIMULATED_ANNEALING.
Increased per-step optimization attempts (times) to scale with CPU count (availableProcessors()/2).
Added JUnit benchmarks/regression tests plus programmatic test-image generation; also added PNGs under test resources.

Reviewed changes

Copilot reviewed 5 out of 10 changed files in this pull request and generated 7 comments.

Show a summary per file

File	Description
`src/main/java/com/bobrust/generator/HillClimbGenerator.java`	Adds SA implementation + temperature schedule helpers; keeps classic variant and dispatches via feature flag.
`src/main/java/com/bobrust/generator/Model.java`	Scales `times` with CPU count to run more optimization chains per step.
`src/main/java/com/bobrust/util/data/AppConstants.java`	Adds `USE_SIMULATED_ANNEALING` feature flag (default `true`).
`src/test/java/com/bobrust/generator/SimulatedAnnealingBenchmark.java`	Adds tests comparing SA vs classic and validating temperature schedule.
`src/test/java/com/bobrust/generator/TestImageGenerator.java`	Adds deterministic in-memory test image generation + a helper `main` to write PNGs.
`src/test/resources/test-images/solid.png`	Adds generated test image (currently unused by tests).
`src/test/resources/test-images/gradient.png`	Adds generated test image (currently unused by tests).
`src/test/resources/test-images/edges.png`	Adds generated test image (currently unused by tests).
`src/test/resources/test-images/photo_detail.png`	Adds generated test image (currently unused by tests).
`src/test/resources/test-images/nature.png`	Adds generated test image (currently unused by tests).

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

Copilot · 2026-04-04T21:11:23Z

src/main/java/com/bobrust/generator/HillClimbGenerator.java

+	 */
+	static float computeCoolingRate(float initialTemp, int maxAge) {
+		int totalIterations = maxAge * 10;
+		float finalTemp = 0.001f;


computeCoolingRate can divide by zero when maxAge is 0 (since totalIterations = maxAge * 10 becomes 0, and the later Math.pow(..., 1.0 / totalIterations) becomes undefined). Add an explicit guard for maxAge <= 0 (and/or totalIterations <= 0) to return a sane rate (e.g., 1.0f) to keep the method safe for all callers.

Suggested change

float finalTemp = 0.001f;

float finalTemp = 0.001f;

if (maxAge <= 0 || totalIterations <= 0) {

return 1.0f; // no cooling when iteration budget is invalid

}

Copilot · 2026-04-04T21:11:23Z

src/main/java/com/bobrust/generator/HillClimbGenerator.java

+		// -1/ln(0.6) ≈ 1.957, but we use avgDelta / 0.5108 which is equivalent
+		return (float) (avgDelta / 0.5108);


estimateTemperature uses the magic constant 0.5108 (approx -ln(0.6)) which makes the math harder to audit and invites drift from the comment. Consider computing it from Math.log(0.6) (or declaring a named constant) to keep the intent precise and self-documenting.

Suggested change

// -1/ln(0.6) ≈ 1.957, but we use avgDelta / 0.5108 which is equivalent

return (float) (avgDelta / 0.5108);

return (float) (avgDelta / -Math.log(0.6));

Copilot · 2026-04-04T21:11:23Z

src/main/java/com/bobrust/generator/Model.java

 	private static final int max_random_states = 1000;
 	private static final int age = 100;
-	private static final int times = 1;
+	private static final int times = Math.max(1, Runtime.getRuntime().availableProcessors() / 2);


Deriving times directly from availableProcessors()/2 can scale work linearly with core count (each processStep() does times full hillclimb runs), which can cause large latency spikes on high-core machines. Consider adding an upper cap and/or making this configurable (e.g., via a constant/setting) so performance remains predictable.

Suggested change

private static final int times = Math.max(1, Runtime.getRuntime().availableProcessors() / 2);

private static final int max_times = 8;

private static final int times = Math.min(max_times,

Math.max(1, Runtime.getRuntime().availableProcessors() / 2));

Copilot · 2026-04-04T21:11:23Z

src/test/java/com/bobrust/generator/SimulatedAnnealingBenchmark.java

+	/** Reflective helper to get the worker from Model (package-private field) */
+	private static Worker getWorker(Model model) {
+		try {
+			var field = Model.class.getDeclaredField("worker");
+			field.setAccessible(true);
+			return (Worker) field.get(model);
+		} catch (Exception e) {
+			throw new RuntimeException(e);
+		}


The test uses reflection to access private internals (Model.worker and Model.addShape), which is brittle and can break on refactors or stricter JVM access rules. Prefer exposing a package-private test hook (e.g., a package-private accessor/add method) or restructuring the benchmark to avoid reflective access.

Copilot · 2026-04-04T21:11:24Z

src/test/java/com/bobrust/generator/SimulatedAnnealingBenchmark.java

+		System.out.println("Hill climb score: " + hillClimbScore);
+		System.out.println("SA score: " + saScore);
+		float improvement = (hillClimbScore - saScore) / hillClimbScore * 100;
+		System.out.println("SA improvement: " + improvement + "%");


Avoid System.out.println in unit tests; it adds noise to CI logs and can slow execution. Prefer assertions-only, or log via the test framework when debugging is explicitly enabled.

Copilot · 2026-04-04T21:11:24Z

src/test/java/com/bobrust/generator/SimulatedAnnealingBenchmark.java

+import org.junit.jupiter.api.BeforeAll;
+import org.junit.jupiter.api.Test;
+
+import java.awt.*;
+import java.awt.image.BufferedImage;
+import java.awt.image.DataBufferInt;
+import java.io.File;
+import java.io.IOException;
+import java.util.ArrayList;
+import java.util.Arrays;
+import java.util.List;
+import javax.imageio.ImageIO;


Several imports appear unused in this test class (BeforeAll, DataBufferInt, File, IOException, ImageIO). Removing unused imports will keep the test file focused and avoid confusion about intended resource-based testing.

Suggested change

import org.junit.jupiter.api.BeforeAll;

import org.junit.jupiter.api.Test;

import java.awt.*;

import java.awt.image.BufferedImage;

import java.awt.image.DataBufferInt;

import java.io.File;

import java.io.IOException;

import java.util.ArrayList;

import java.util.Arrays;

import java.util.List;

import javax.imageio.ImageIO;

import org.junit.jupiter.api.Test;

import java.awt.*;

import java.awt.image.BufferedImage;

import java.util.ArrayList;

import java.util.Arrays;

import java.util.List;

Copilot · 2026-04-04T21:11:24Z

src/test/java/com/bobrust/generator/SimulatedAnnealingBenchmark.java

+		// SA should not be dramatically worse (allow 5% tolerance due to stochastic nature)
+		assertTrue(saScore <= hillClimbScore * 1.05f,
+			"SA score (" + saScore + ") should not be significantly worse than hill climb (" + hillClimbScore + ")");
+	}
+
+	// ---- Test 4: Regression — No Worse Than Baseline on Multiple Images ----
+
+	@Test
+	void testSANeverSignificantlyWorse() {
+		BufferedImage[] images = {
+			TestImageGenerator.createSolid(),
+			TestImageGenerator.createGradient(),
+			TestImageGenerator.createEdges(),
+		};
+		String[] names = {"solid", "gradient", "edges"};
+		int maxShapes = 30; // Small for test speed
+
+		for (int idx = 0; idx < images.length; idx++) {
+			float hcScore = runGenerator(images[idx], maxShapes, false);
+			float saScore = runGenerator(images[idx], maxShapes, true);
+			System.out.println(names[idx] + " — HC: " + hcScore + ", SA: " + saScore);
+
+			// SA should never be more than 5% worse (stochastic tolerance)
+			assertTrue(saScore <= hcScore * 1.05f,
+				names[idx] + ": SA (" + saScore + ") should not be significantly worse than HC (" + hcScore + ")");


These assertions compare two stochastic optimizers using unseeded randomness (Worker.getRandom() uses ThreadLocalRandom), so the 5% tolerance can still lead to intermittent CI failures. To avoid flaky tests, consider injecting a deterministic RNG/seed for tests or running multiple trials and asserting on an aggregate (median/mean) rather than a single run.

Copilot AI review requested due to automatic review settings April 4, 2026 21:06

Copilot started reviewing on behalf of VoX April 4, 2026 21:07 View session

VoX merged commit 88e4f8a into main Apr 4, 2026
2 checks passed

Copilot AI reviewed Apr 4, 2026

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Feature: Simulated annealing optimization#1

Feature: Simulated annealing optimization#1
VoX merged 1 commit intomainfrom
feature/simulated-annealing

VoX commented Apr 4, 2026

Uh oh!

Uh oh!

Copilot AI left a comment

Uh oh!

Copilot AI Apr 4, 2026

Uh oh!

Copilot AI Apr 4, 2026

Uh oh!

Copilot AI Apr 4, 2026

Uh oh!

Copilot AI Apr 4, 2026

Uh oh!

Copilot AI Apr 4, 2026

Uh oh!

Copilot AI Apr 4, 2026

Uh oh!

Copilot AI Apr 4, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

		// -1/ln(0.6) ≈ 1.957, but we use avgDelta / 0.5108 which is equivalent
		return (float) (avgDelta / 0.5108);

	// -1/ln(0.6) ≈ 1.957, but we use avgDelta / 0.5108 which is equivalent
	return (float) (avgDelta / 0.5108);
	return (float) (avgDelta / -Math.log(0.6));

-	private static final int times = Math.max(1, Runtime.getRuntime().availableProcessors() / 2);
+	private static final int max_times = 8;
+	private static final int times = Math.min(max_times,
+			Math.max(1, Runtime.getRuntime().availableProcessors() / 2));

Conversation

VoX commented Apr 4, 2026

Summary

Changes

How SA works

Test plan

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Copilot AI Apr 4, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 4, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 4, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 4, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 4, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 4, 2026

Choose a reason for hiding this comment

Uh oh!

Copilot AI Apr 4, 2026

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants