One update, term by term
What a single IDBD update does
sample 0
A three-input learner on the task y = 2 x0 - x1 + 0 x2. Advance one sample and read the update in the order the algorithm performs it. Every number below is read from the Rust core before and after its own update.
The sample
| Input | Value | True coefficient |
|---|
Follow one input
State after this sample
| Input | Weight | Alpha | Trace |
|---|
The update for x0, in order
Twenty inputs, one shared stream
The experiment
sample 0
Paused
Changing any setting above starts a new run from sample 0.
Run definition: task, seed, settings and versions
Learned weight per input
Select an input to inspect it. Arrow keys move between inputs.
sample 0
Fills in after the first sample. Choose Step or Run.
Prediction error over samples
sample 0
Fills in after the first sample. Choose Step or Run.
The chart keeps the most recent 2,000 samples. A run starts all three learners at zero weights, so its first samples include an initial transient. This single run on synthetic data does not show that any one rule is better in general.
IDBD step sizes, all 20 inputs
Alpha is the size of the learning adjustment an input receives, not its importance or a confidence.
sample 0
Fills in after the first sample. Choose Step or Run.
World events
Changes to the target-generating relationship, applied before a sample.
sample 0
Fills in after the first sample. Choose Step or Run.
Input x0
The true coefficient is known simulator information, not something any learner is told.
sample 0
Fills in after the first sample. Choose Step or Run.
Values
The same state the charts draw, as numbers rounded for reading. Export run saves every value at full precision.
sample 0
Fills in after the first sample. Choose Step or Run.
| Input | Value | True coefficient | IDBD weight | LMS weight | Alpha | Trace | Autostep weight | Autostep alpha |
|---|