Google

When multiple evaluation metrics matter, such as accuracy and running time of a learning algorithm, one way to combine them is to define an acceptable value for one metric and then optimize another metric subject to meeting that criterion. In the passage's example, running time is the satisficing metric and accuracy is the optimizing metric.

Optimizing and Satisficing Metrics

Machine Learning Yearning, a free ebook from Andrew Ng, teaches you how to structure Machine Learning projects.
https://www.deeplearning.ai/machine-learning-yearning/

Note: The content of the book is aligned with the Coursera Deeplearning.ai specialization.  https://www.deeplearning.ai/deep-learning-specialization/ 

Machine Learning Yearning (Deeplearning.ai)

A satisficing metric only has to be good enough by meeting a chosen value. For example, if running in 100ms is acceptable, then running time can be used as a satisficing metric that should take at most 100ms.

Satisficing Metric

An optimizing metric is the evaluation metric to optimize after the satisficing metric criteria have been set. In the passage's classifier example, accuracy is the optimizing metric when accuracy is maximized subject to meeting the running time criterion.

Optimizing Metric

When trading off N different criteria, such as a model's binary file size, running time, and accuracy, one approach is to set N-1 criteria as satisficing metrics that must meet certain values, and define the final criterion as the optimizing metric. For example, one can set acceptable thresholds for binary file size and running time, then optimize accuracy given those constraints.

Choosing N-1 Satisficing Metrics and One Optimizing Metric

For a hardware device that uses a microphone to listen for a particular wakeword, both false positive rate and false negative rate matter. One reasonable performance goal is to minimize the false negative rate as the optimizing metric, subject to no more than one false positive every 24 hours of operation as the satisficing metric.

Wake Word Detection as Optimizing and Satisficing Example

Once a team is aligned on the evaluation metric to optimize, they will be able to make faster progress.

Team Alignment on Evaluation Metric Speeds Progress

Identifying Optimizing vs. Satisficing Metrics

Definition of Satisficing Metrics

Combining Multiple Evaluation _____ Metrics

Roles of Evaluation Metrics

Applying Satisficing and Optimizing Metrics

Question:
Analyze the difference between an optimizing metric and a satisficing metric in the context of evaluating a machine learning algorithm. Provide an example to illustrate how they are combined.

Sample answer:
When evaluating an algorithm with multiple goals, an optimizing metric is the primary performance measure that you actively seek to maximize or minimize. In contrast, a satisficing metric acts as a constraint or threshold; the algorithm only needs to be 'good enough' to meet an acceptable value on this metric. For example, if you care about accuracy and running time, you might define running time as a satisficing metric that must be under 100ms. Once that criterion is met, you optimize accuracy, meaning you select the model with the highest accuracy that still runs in under 100ms.

Key points:
- Optimizing metric is maximized or minimized.
- Satisficing metric only needs to meet an acceptable, 'good enough' threshold.
- They are combined by maximizing the optimizing metric subject to the satisficing metric's criteria.
- Example provided (e.g., maximizing accuracy while keeping running time under 100ms).

Rubric:
A strong answer should clearly distinguish between maximizing a metric (optimizing) and merely meeting an acceptable threshold for a metric (satisficing), and accurately relate this to a combined evaluation strategy using a practical example.

Explaining Satisficing vs. Optimizing

Case context:
You are building a wake word detection system for a smart speaker. You care about the model's accuracy, but the device strictly has only 50MB of RAM available to load your model into memory.

Question:
Based on the principles of optimizing and satisficing metrics, how should you structure your evaluation metrics to decide which model is best to deploy? Specify which metric is optimizing and which is satisficing.

Sample answer:
In this scenario, the memory footprint must be treated as a satisficing metric. Because the device strictly has 50MB of RAM, any model requiring more than 50MB is completely unacceptable, while any model under 50MB is 'good enough' to fit on the device. Accuracy, on the other hand, should be treated as the optimizing metric. Therefore, the evaluation strategy should be to maximize accuracy, subject to the classifier meeting the satisficing criterion of using 50MB of RAM or less.

Key points:
- Memory footprint/RAM is the satisficing metric.
- The satisficing criterion is set at 50MB or less.
- Accuracy is the optimizing metric.
- The goal is to maximize accuracy subject to the RAM constraint.

Rubric:
A correct response will correctly identify RAM/memory footprint as the satisficing metric due to the strict threshold, and accuracy as the optimizing metric, demonstrating the relationship of maximizing one subject to the other.

Diagnosing Algorithm Performance Constraints

Question:
In the passage's example, why is running time designated as a 'satisficing metric' rather than an optimizing metric?

Sample answer:
Running time is designated as a satisficing metric because the classifier does not need to be infinitely fast; it just needs to be 'good enough' to meet an acceptable running time criterion, such as taking at most 100ms.

Key points:
- It just has to be 'good enough'.
- It involves meeting an acceptable running time criterion (e.g., at most 100ms).

Rubric:
The answer should mention that running time only requires meeting an 'acceptable' threshold or being 'good enough' (e.g., at most 100ms), rather than needing to be maximized or minimized indefinitely.

Learn Before

Related