Fix bestdose posterior by R-garreau · Pull Request #281 · LAPKB/PMcore

R-garreau · 2026-05-20T20:25:36Z

This pull request introduces a significant enhancement to the BestDose algorithm by adding support for configurable Stage 2 optimization strategies. It allows users to choose between the original "dual" optimization (posterior and uniform) and a new "posterior-only" strategy, improving flexibility and transparency. The changes include API extensions, a new optimization pathway, and improved output for support points. Additionally, there are bug fixes and refactoring for clarity and correctness.

Major new features and API changes:

Added the OptimizationStrategy enum (with Dual and PosteriorOnly variants) to control Stage 2 optimization in BestDoseProblem, with Dual as the default for backward compatibility. Methods to set and query the strategy were added to BestDoseProblem, and the API was updated to expose this option. (src/bestdose/types.rs, src/bestdose/mod.rs, examples/bestdose_posterior.rs, [[1]](https://github.com/LAPKB/PMcore/pull/281/files#diff-5579cab66369604239b0d08fb7a22468dc271026dbceefc5999acc66b616fee0R186-R212), [[2]](https://github.com/LAPKB/PMcore/pull/281/files#diff-63aea7f46152f96d93294845ba544f57e0386115a6a4ac1adaecd9795be08fd7L305-R305), [[3]](https://github.com/LAPKB/PMcore/pull/281/files#diff-63aea7f46152f96d93294845ba544f57e0386115a6a4ac1adaecd9795be08fd7R802-R807), [[4]](https://github.com/LAPKB/PMcore/pull/281/files#diff-63aea7f46152f96d93294845ba544f57e0386115a6a4ac1adaecd9795be08fd7R833-R837))
The optimize method of BestDoseProblem now dispatches to either the dual or posterior-only optimization path based on the selected strategy, with appropriate logging and result handling. (src/bestdose/mod.rs, [src/bestdose/mod.rsR772-R790](https://github.com/LAPKB/PMcore/pull/281/files#diff-63aea7f46152f96d93294845ba544f57e0386115a6a4ac1adaecd9795be08fd7R772-R790))
New example examples/bestdose_posterior.rs demonstrates the use of the posterior-only optimization strategy and prints detailed summaries of posterior support points and weights. (examples/bestdose_posterior.rs, [examples/bestdose_posterior.rsR1-R161](https://github.com/LAPKB/PMcore/pull/281/files#diff-dba6e21e48e516e56e39cd8bc1796d949ad17074b73a261ecda029b04ffe2e38R1-R161))

Algorithmic and implementation improvements:

Refactored the Stage 2 and Stage 3 optimization logic: introduced a shared finalize_optimization function, and added a dedicated posterior_optimization function for the posterior-only pathway. The dual optimization path now delegates to this shared logic. (src/bestdose/optimization.rs, [[1]](https://github.com/LAPKB/PMcore/pull/281/files#diff-6a560c2551efb495e2b115bb86555df9abd5ba9b0f1159ea51ff5d68391d7171R183-R254), [[2]](https://github.com/LAPKB/PMcore/pull/281/files#diff-6a560c2551efb495e2b115bb86555df9abd5ba9b0f1159ea51ff5d68391d7171L258-R330))
Improved the NPAGFULL11 filtering in posterior calculation to use direct Bayesian updating rather than a separate optimization step, clarifying the intent and improving transparency. (src/bestdose/posterior.rs, [[1]](https://github.com/LAPKB/PMcore/pull/281/files#diff-0ddd7fe773579ff845c545728699cd054a3823fd4bed59b7d8de18d2d08f2a35L89-R89), [[2]](https://github.com/LAPKB/PMcore/pull/281/files#diff-0ddd7fe773579ff845c545728699cd054a3823fd4bed59b7d8de18d2d08f2a35L104-R178))

Bug fixes:

Fixed a bug in the cost calculation where population means were incorrectly weighted by posterior weights instead of prior weights. (src/bestdose/cost.rs, [src/bestdose/cost.rsL513-R513](https://github.com/LAPKB/PMcore/pull/281/files#diff-072b6f7fdcffee448fdc9b9051fa53f7ea2d449daea95b09fd932c2f09799ddcL513-R513))

Minor changes and improvements:

Enhanced output in examples to print support point summaries and improved logging for optimization strategy selection. (examples/bestdose.rs, [[1]](https://github.com/LAPKB/PMcore/pull/281/files#diff-c762aa46050859e4c42d5576d0408956abd81bbbc622b9ec44e25378f31bf724R134-R166), src/bestdose/mod.rs, [[2]](https://github.com/LAPKB/PMcore/pull/281/files#diff-63aea7f46152f96d93294845ba544f57e0386115a6a4ac1adaecd9795be08fd7R713))
Minor code cleanup and doc updates. (src/bestdose/posterior.rs, [[1]](https://github.com/LAPKB/PMcore/pull/281/files#diff-0ddd7fe773579ff845c545728699cd054a3823fd4bed59b7d8de18d2d08f2a35L56), [[2]](https://github.com/LAPKB/PMcore/pull/281/files#diff-0ddd7fe773579ff845c545728699cd054a3823fd4bed59b7d8de18d2d08f2a35L89-R89))

Other:

Updated example data file references for consistency. (examples/bestdose.rs, [examples/bestdose.rsL76-R76](https://github.com/LAPKB/PMcore/pull/281/files#diff-c762aa46050859e4c42d5576d0408956abd81bbbc622b9ec44e25378f31bf724L76-R76))

These changes provide a more flexible and robust framework for dose optimization, enabling easier experimentation and clearer results.

- remove burke and implement direct bayes weight update - update bimodal_ke example

- split methods of dual optimization - User can now chose to run posterior only or dual (posterior and uniform) - add examples (`bestdose_posterior.rs`) to compare dual and posterior only Co-Authored-By: Julian Otalvaro <1023006+Siel@users.noreply.github.com> Co-Authored-By: Markus Hovd <66058642+mhovd@users.noreply.github.com>

Copilot

Pull request overview

This PR extends the BestDose API and optimization pipeline to support a configurable Stage 2 strategy (dual vs posterior-only), refactors Stage 3 finalization to be shared, and updates posterior filtering to use a direct Bayes update.

Changes:

Added OptimizationStrategy and plumbing on BestDoseProblem to select Stage 2 optimization behavior (Dual default vs PosteriorOnly).
Refactored Stage 2/3 optimization code by introducing a shared finalize_optimization and a dedicated posterior-only optimization path.
Reworked NPAGFULL11 filtering to compute posterior weights via direct Bayesian updating, and updated examples to show support-point summaries and posterior-only usage.

Reviewed changes

Copilot reviewed 7 out of 7 changed files in this pull request and generated 2 comments.

Show a summary per file

File	Description
`src/bestdose/types.rs`	Introduces `OptimizationStrategy` and stores it on `BestDoseProblem`.
`src/bestdose/mod.rs`	Re-exports the new enum, adds setters/getters, and dispatches `optimize()` based on the selected strategy.
`src/bestdose/optimization.rs`	Adds posterior-only Stage 2 path and factors Stage 3 result building into `finalize_optimization()`.
`src/bestdose/posterior.rs`	Replaces the prior burke-based step with a direct Bayes update for NPAGFULL11 filtering.
`src/bestdose/cost.rs`	Adjusts weight usage in cost computation (but currently reintroduces a weighting bug; see comment).
`examples/bestdose.rs`	Updates prior file path and prints support-point/weight summaries.
`examples/bestdose_posterior.rs`	New example demonstrating posterior-only optimization and printing detailed summaries.

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

R-garreau · 2026-05-21T10:04:55Z

            sumsq_i += se;
            // Calculate population mean using PRIOR probabilities
-            y_bar[j] += prior_prob * pj;
+            y_bar[j] += post_prob * pj;


not an error.
the bias trade-off is calculated based on the posterior, just need to update documentation

mhovd · 2026-05-21T06:54:42Z

        tracing::info!("  Support points: {}", posterior_theta.matrix().nrows());
        tracing::info!("  Target type: {:?}", target_type);
        tracing::info!("  Bias weight (λ): {}", bias_weight);
+        tracing::info!("  Optimization strategy: {}", OptimizationStrategy::Dual);

        Ok(BestDoseProblem {


This is a good catch

mhovd · 2026-05-21T06:55:21Z

+            settings,
+            doserange,
+            bias_weight,
+            optimization_strategy: OptimizationStrategy::Dual,


Should this be hardcoded?

mhovd · 2026-05-21T06:55:46Z

+        match self.optimization_strategy {
+            OptimizationStrategy::Dual => {
+                tracing::info!("╔══════════════════════════════════════════════════════════╗");
+                tracing::info!("║            BestDose Algorithm: STAGE 2 & 3               ║");
+                tracing::info!("║        Dual Optimization + Final Predictions             ║");
+                tracing::info!("╚══════════════════════════════════════════════════════════╝");
+
+                optimization::dual_optimization(&self)
+            }
+            OptimizationStrategy::PosteriorOnly => {
+                tracing::info!("╔══════════════════════════════════════════════════════════╗");
+                tracing::info!("║            BestDose Algorithm: STAGE 2 & 3               ║");
+                tracing::info!("║    Posterior Optimization Only + Final Predictions       ║");
+                tracing::info!("╚══════════════════════════════════════════════════════════╝");


While pretty, perhaps we should stick to one-line logs?

My take on this is do we need it ?
While this appear in console when running the inline command, it will never appear in the app.
Plus, when logging the run to find why it crashes, it makes the log unecessary long in my opinion.

i'd rather have a count that could be use to track progress.

Agree. And if this is intented for console, then we should not print it to logs using tracing, but use println! instead. But for a binary, I think it should be limited.

mhovd · 2026-05-21T06:58:58Z

+impl Default for OptimizationStrategy {
+    fn default() -> Self {
+        Self::Dual
+    }
+}


I think an argument can be made to have the PosteriorOnly (perhaps renamed to Posterior or more explicitly MAP, as the default.

mhovd · 2026-05-21T07:06:01Z

@@ -510,7 +510,7 @@ pub fn calculate_cost(problem: &BestDoseProblem, candidate_doses: &[f64]) -> Res
            let se = (obs_val - pj).powi(2);
            sumsq_i += se;
            // Calculate population mean using PRIOR probabilities


Suggested change

// Calculate population mean using PRIOR probabilities

// Calculate population mean using the posterior probabilities

mhovd · 2026-05-21T07:08:37Z

+        if prior_weight <= 0.0 {
+            continue;
+        }


This seems unnecessary

mhovd · 2026-05-21T07:09:17Z

+        }
+
+        let mut log_weight = prior_weight.ln();
+        let mut is_zero = false;


To improve code clarity, it is better to have this in the if statement itself, since it is so small.

mhovd · 2026-05-21T07:12:17Z

+pub enum OptimizationStrategy {
+    Dual,
+    PosteriorOnly,
+}
+


Could we perhaps "rebrand" the Dual and PosteriorOnly to be more informative?

Siel and others added 5 commits March 24, 2026 20:44

fix(bestdose) variance is calculated using the post_prob

9fba0a4

fix(bestdose) variance is calculated using the post_prob

ef44694

Merge branch 'main' into fix/bestdose

7603584

fix : implement direct bayes weight update

96eb6c7

- remove burke and implement direct bayes weight update - update bimodal_ke example

Copilot AI review requested due to automatic review settings May 20, 2026 20:25

Copilot started reviewing on behalf of R-garreau May 20, 2026 20:26 View session

Copilot AI reviewed May 20, 2026

View reviewed changes

mhovd reviewed May 21, 2026

View reviewed changes

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Fix bestdose posterior#281

Fix bestdose posterior#281
R-garreau wants to merge 5 commits into
mainfrom
fix-bestdose-posterior

R-garreau commented May 20, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

R-garreau May 21, 2026

Uh oh!

mhovd May 21, 2026

Uh oh!

mhovd May 21, 2026

Uh oh!

mhovd May 21, 2026

Uh oh!

R-garreau May 21, 2026

Uh oh!

mhovd May 21, 2026

Uh oh!

mhovd May 21, 2026

Uh oh!

mhovd May 21, 2026

Uh oh!

mhovd May 21, 2026

Uh oh!

mhovd May 21, 2026

Uh oh!

mhovd May 21, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

	// Calculate population mean using PRIOR probabilities
	// Calculate population mean using the posterior probabilities

Conversation

R-garreau commented May 20, 2026

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants