8 posts tagged with "Experiments"

See how each completed trial moves through its parameters and targets in a single parallel coordinates chart.

---

What's included​

Parallel coordinates​

• Open Parallel coordinates from an experiment's Visualization page
• Each completed trial is drawn as one line, running across every parameter on the left and the selected targets on the right
• Every axis has its own scale, so parameters and targets with different units can be compared side by side
• Categorical parameters show one tick per category
• Hover a line to see that trial's values on every axis

Reorder axes​

• Drag an axis title left or right to move that axis
• Place related parameters next to each other, or move a parameter next to a target to see how they relate

Color by target​

• Lines are colored by the last selected target, so you can follow where trials end up
• Drag any target from the sidebar onto Color to color by it instead
• A target dropped onto Color is added to the chart if it is not already selected
• Use the color scale slider to show only trials within a range of target values
• Clear Color to show every line in a single color

Overlapping trials​

• Trials that share the same value on an axis are spread slightly apart, so each one stays visible
• Trials with unique values stay exactly on their tick

---

Notes​

• Only completed trials with a value on every axis are shown
• Axis order and color choice reset when you leave the Parallel coordinates tab

Weight each discrete parameter with its own coefficient, and see a preview of the rule as you build it.

---

What's included​

Linear replaces Sum​

• The Sum constraint for discrete numerical parameters is now called Linear
• Give each parameter a coefficient, for example 2 × x0 + x1 <= 100
• Coefficients start at 1, which gives a plain sum
• Use a negative coefficient to compare two parameters

Preview of your rule​

• A preview below the form shows the rule as you fill it in, for example 2 · x0 + x1 <= 100

Example: keep x1 at or below x0​

To make sure x1 ≤ x0, write it as x0 − x1 ≥ 0 and add a Linear constraint:

• Coefficients: x0 = 1, x1 = −1
• Operator: >=
• Right-Hand Side: 0

Other improvements​

• Negative numbers and decimals can be typed directly in Linear and Product constraints
• Removing a parameter from the Selected list now also removes its coefficient or condition

---

Notes​

• Your existing Sum constraints are now Linear constraints with every coefficient set to 1, so they work exactly as before
• Tolerance only applies to the = and != operators

Follow optimization progress more clearly with completed results ordered as observations and predictions grouped separately.

---

What's included​

A clearer optimization sequence​

• Completed target results are numbered in completion order, without gaps from trials that did not produce a measurement
• The original trial number and completion time remain available in the point details
• Predictions for Pending and Accepted trials appear after the completed observations
• A dashed divider separates measured results from predictions across selected targets

---

Notes​

• The observation number can differ from the original trial number when trials finish out of order or do not produce a target result
• Predictions remain model estimates and include error bars showing one posterior standard deviation above and below the mean

Explore how each parameter influences experiment targets with SHAP-based importance visualizations.

---

What's included​

Parameter importance​

• Open Parameter importance from an experiment's Visualization page
• Select a target to inspect how its parameters contribute to the surrogate model output
• Compare parameters using their mean absolute SHAP values
• See the direction and distribution of each parameter's impact in the beeswarm plot
• Use the feature-value scale to focus on observations with low or high numeric parameter values

---

Notes​

• Parameter importance requires enough completed measurements to fit the surrogate model
• SHAP values describe model behavior and do not imply that a parameter causes a target outcome

View BayBE target predictions and their uncertainty for Pending and Accepted trials.

---

What's included​

Predictions in trial cards​

• Pending and Accepted trial cards show each target's predicted mean and posterior standard deviation as mean ± std
• Predictions update as completed trial results are added to the experiment

Predictions in Convergence​

• Prediction points appear alongside completed measurements on the same target axis
• Error bars extend from mean − std to mean + std to show model uncertainty
• Hover over a prediction to inspect its value and error-bar bounds
• Use Error bar opacity (%) to adjust the visibility of the error bars, including hiding them at 0%

---

Notes​

• Predictions are model estimates and are not measured target results
• Error-bar bounds are not observed minimum and maximum values or a confidence interval

Sync your experiment trials to a linked Foundry dataset with one click, writing a snapshot of all trial data as a Parquet file.

---

What's included​

Sync button on the trials page​

A new Sync button sits in the top-right toolbar of the experiment trials page, next to the Add and Upload controls.

• Click Sync to export all trials for the experiment to the linked Foundry dataset
• The button shows a loading spinner while the sync runs and toasts the number of rows synced on completion
• The button is disabled when there are no trials or no Foundry dataset is linked to the experiment

What gets synced​

• Only trials where all parameters are complete are included, regardless of state
• Each sync overwrites the previous snapshot

---

Notes​

• The Sync button requires a Foundry dataset to be linked to the experiment at creation time
• Each sync is a full snapshot — incremental/append writes are not currently supported

New desirability convergence chart for tracking optimization progress across experiment trials.

---

What's included​

How it works​

• Select a numerical parameter from the parameter sidebar to color data points by that parameter's value
• A continuous color scale shows how parameter values relate to desirability scores across trials
• The best so far line tracks the highest desirability achieved up to each trial

• Select a categorical parameter to color data points by category
• A discrete color legend identifies each category value

---

Notes​

• The convergence chart is available under the Desirability tab in the Visualization section
• Chart interactions and layout may continue to evolve during the beta phase
• If you notice any unexpected behavior, please share steps to reproduce

Split convergence plot into a multi-panel grid when multiple targets are selected.

---

What's included​

How it works​

• Select multiple targets from the target sidebar
• The convergence plot automatically splits into a grid — one panel per target
• Each panel shares the same x-axis (iteration index) but has an independent y-axis scale

• Click a data point to highlight the corresponding trial across all panels

• Shift-drag to draw a brush selection across panels
• Points outside the selection are dimmed
• Up to 6 targets can be displayed at a time

---

Notes​

• Single target selection continues to show the standard convergence plot
• Trellis layout and interactions may continue to evolve during the beta phase
• If you notice any unexpected behavior, please share steps to reproduce