| Princeton University Technical Report | 1994 |
Abstract: Hidden Markov models (HMMs) and other time series models assign probabilities to long sequences of events. Avoiding underflow is arguably the central difficulty in calculating the probability of such sequences. This technical report presents an elegant and efficient C library for representing and manipulating probability values without underflow. Use of the library results in simpler code whose execution time compares favorably with traditional numerical methods. Thus, the library provides a strong foundation on which to build large stochastic modeling systems. Our abstraction also suggests a natural extension of the IEEE 754 floating point standard, to better support statistical computation.