Table of Contents

Class LearningCurveFitDetectorOptions

Namespace
AiDotNet.Models.Options
Assembly
AiDotNet.dll

Configuration options for the Learning Curve Fit Detector, which analyzes training progress to determine when a model has converged or is unlikely to improve further.

public class LearningCurveFitDetectorOptions
Inheritance
LearningCurveFitDetectorOptions
Inherited Members

Remarks

The Learning Curve Fit Detector monitors the training progress of a machine learning model by analyzing the pattern of error reduction over time. It fits a mathematical curve to the error values and uses this to predict whether continued training is likely to yield significant improvements. This can help automatically determine when to stop training.

For Beginners: When training a machine learning model, it's important to know when to stop. Training for too long can waste time or cause overfitting (where the model performs well on training data but poorly on new data), while stopping too early might leave the model under-trained.

The Learning Curve Fit Detector is like a smart assistant that watches how your model's performance improves during training. It looks at the pattern of improvement and tries to predict whether continuing to train will give meaningful benefits or if the model has already learned as much as it can from the data.

Think of it like watching someone learn a new skill:

  • At first, they improve quickly (steep learning curve)
  • Over time, the rate of improvement slows down (flattening curve)
  • Eventually, they reach a plateau where more practice yields minimal improvement

This class lets you configure how the detector decides when that plateau has been reached, allowing you to automatically stop training at the right time.

Properties

ConvergenceThreshold

Gets or sets the threshold that determines when the model is considered to have converged.

public double ConvergenceThreshold { get; set; }

Property Value

double

The convergence threshold, defaulting to 0.01 (1%).

Remarks

This threshold represents the predicted percentage improvement in the error metric if training were to continue. When the predicted improvement falls below this threshold, the detector considers the model to have converged. Lower values require more evidence of convergence before stopping training.

For Beginners: This setting controls how strict the detector is when deciding if your model has finished learning.

The convergence threshold represents how much improvement we expect to see if training continues. For example, with the default value of 0.01 (or 1%):

  • If the detector predicts that continuing to train would improve your model's performance by less than 1%, it will suggest stopping because further training isn't worth the time.
  • If it predicts an improvement of more than 1%, it will suggest continuing training.

You can adjust this value based on your needs:

  • Lower values (like 0.001 or 0.1%) make the detector more patient, continuing training until even tiny improvements are unlikely. This might give slightly better results but takes longer.
  • Higher values (like 0.05 or 5%) make the detector stop training earlier, saving time but potentially leaving some performance on the table.

The default of 1% is a good balance for most problems, ensuring significant improvements are captured while avoiding diminishing returns.

MinDataPoints

Gets or sets the minimum number of data points (training iterations) required before the detector will attempt to predict convergence.

public int MinDataPoints { get; set; }

Property Value

int

The minimum number of data points, defaulting to 5.

Remarks

To make reliable predictions about convergence, the detector needs a sufficient history of error values from previous training iterations. This parameter specifies how many data points must be collected before the detector will begin making predictions. Setting this too low may result in premature convergence detection, while setting it too high delays detection.

For Beginners: This setting determines how many training steps must complete before the detector starts making decisions about stopping.

To predict whether your model has converged, the detector needs to see a pattern in how the performance changes over time. This setting specifies how many measurements it needs before it starts making these predictions.

With the default value of 5:

  • During the first 5 training iterations, the detector will always recommend continuing because it doesn't have enough information yet
  • After 5 iterations, it will start analyzing the pattern to decide if training should continue

Think of it like trying to predict the weather:

  • With only 1-2 measurements, it's hard to see a reliable pattern
  • With 5+ measurements, you can start to see trends and make better predictions

You might want to increase this value (to 10 or more) for complex models or noisy training processes where performance fluctuates a lot between iterations. For simpler, more stable training processes, the default of 5 is usually sufficient.

Edit this page