Occam’s razor principle:
Having two hypotheses (here, decision boundaries) that has the same empirical risk (here, training error), a short explanation (here, a boundary with fewer parameters) tends to be more valid than a long explanation.
In your example, both A and B have zero training error, thus B (shorter explanation) is preferred.
What if training error is not the same?
If boundary A had a smaller training error than B, selecting becomes tricky. We need to quantify "explanation size" the same as "empirical risk" and combine the two in one scoring function, then proceed to compare A and B. An example would be Akaike Information Criterion (AIC) that combines empirical risk (measured with negative log-likelihood) and explanation size (measured with the number of parameters) in one score.
As a side note, AIC cannot be used for all models, there are many alternatives to AIC too.
Relation to validation set
In many practical cases, when model progresses toward more complexity (larger explanation) to reach a lower training error, AIC and the like can be replaced with a validation set (a set on which the model is not trained). We stop the progress when validation error (error of model on validation set) starts to increase. This way, we strike a balance between low training error and short explanation.