The team created computer simulations of fires to not only teach the neural net to predict an imminent flashover but also evaluate its abilities, generating 4,033 training examples, and 504 test cases. The system, nicknamed P-Flash, was at least 80 per cent accurate with these simulations, according to a paper detailing the research that was published in the Proceedings of the AAAI Conference on Artificial Intelligence. Most of the errors were false positives, the researchers said. However, the accuracy dropped to as low as 25 per cent when the algorithm was given temperature sensor data taken from real fires in 13 experiments conducted by the UL Firefighter Safety Research Institute, depending on what room the fire started in.