Lesson 01 — Evidence before trust
Series: index · next: 02 — Assumptions of a model
An evaluation of an AI component is an observation: a reward (what you got)
against a baseline (what you would have got otherwise). The advantage math in
advantage.h turns a list of observations into a mean, a
spread, and a confidence.
The three rules
- Gain is relative. A 92% score means little without the baseline (e.g. a majority-class guess or the previous model).
- Few observations earn little trust.
confidence = n / (n + prior), prior 3 by default: 1 run gives 0.25, 3 runs 0.5, 12 runs 0.8. - No evidence is unknown, not zero. An untested model is flagged, not scored 0.
In XML
<sourceAdvantage id="neural-net-xor" kind="model">
<evidence weight="1.0">
<observation reward="0.97" baseline="0.50"/>
<observation reward="0.95" baseline="0.50"/>
</evidence>
</sourceAdvantage>
(The values above are illustrative, not measurements.)
Exercise
Three runs give gains 0.4, 0.5, 0.45. What is the confidence? Then add nine more similar runs: how does it change? Answer: 3/(3+3)=0.5; 12/(12+3)=0.8.
Try it: cd advantage && make run-analyze.