A Learned Fire Model Fails Where Its Training Fires Never Went
A model fitted statistically to a large archive of past fires can outperform a physics-based one on the conditions that archive contains, because it absorbs the local quirks nobody wrote an equation for. Its weakness is the mirror image of that strength: it interpolates confidently between the fires it has seen and has no mechanism to fall back on outside them, which is precisely the situation on a record-breaking fire. A physics-based model reasons from mechanism, so it extrapolates on principle, but crudely, and it will be beaten on ordinary days. The questions to ask about any learned spread or detection model are what fires are in its training set, whether your case sits inside them, whether it can express that it does not know, and whether its inputs will still exist on the day you need it. The confusion resolved is that a higher reported accuracy is a claim about the test set, not about the next fire. This is a choice between tools, not a study of how such models are built. After this Concept you can weigh a learned model against a physical one for a specific case instead of in general.
This Concept is waiting for its first lesson!
A model fitted statistically to a large archive of past fires can outperform a physics-based one on the conditions that archive contains, because it absorbs the local quirks nobody wrote an equation for. Its weakness is the mirror image of that strength: it interpolates confidently between the fires it has seen and has no mechanism to fall back on outside them, which is precisely the situation on a record-breaking fire. A physics-based model reasons from mechanism, so it extrapolates on principle, but crudely, and it will be beaten on ordinary days. The questions to ask about any learned spread or detection model are what fires are in its training set, whether your case sits inside them, whether it can express that it does not know, and whether its inputs will still exist on the day you need it. The confusion resolved is that a higher reported accuracy is a claim about the test set, not about the next fire. This is a choice between tools, not a study of how such models are built. After this Concept you can weigh a learned model against a physical one for a specific case instead of in general.
Are you a teacher? Sign in to start contributing.
Sign In