Metrics for the Crossing example

[patham9]

We can define two domain-independent metrics to guide future improvement:

Prediction accuracy: number of confirmed predictions / number of predictions
Here, instead of just counting 1 for each confirmed prediction, the temporal distance between when the input occurred and the expected time for the predicted event to occur needs to be added additionally to weight it (to penalize temporal inaccuracy).

Prediction propensity: number of predictions / fixed time interval

Surprise: the predicted event didn't occur, where we will add a parameter that allows to set a temporal difference tolerance somewhere between 0 and anticipation_tolerance, to have a configurable "degree of anomaly"

And I suggest a domain-specific danger measurement, that increases when a car is predicted to be at a certain position close to where a pedestrian is or will be according to predictions.

All this will allow to use NARS for predicting dangerous situations in the crossing as well as to report anomalous cases based on the predictions it is already successfully inferring.

Also: Consider to use an ontology for reasoning about types of anomalies as Enzo suggested, this would generalize these ideas into a bigger picture.