Many epidemiological studies have used proximity to sources as air pollution exposure assessment method. However, proximity measures are not generally good surrogates because of their complex non-linear relationship with exposures. Neuro-fuzzy inference systems (NFIS) can be used to map complex non-linear systems, but its usefulness in exposure assessment has not been extensively explored. We present a novel approach for exposure assessment using NFIS, where the inputs of the model were easily-obtainable proximity measures, and the output was residential exposure to an air pollutant. We applied it to a case-study on NH3 pollution, and compared health effects and exposures estimated from NFIS, with those obtained from emission-dispersion models, and linear and non-linear regression proximity models, using 10-fold cross validation. The agreement between emission-dispersion and NFIS exposures was high (Root-mean-square error (RMSE) =0.275, correlation coefficient (r)=0.91) and resulted in similar health effect estimates. Linear models showed poor performance (RMSE=0.527, r=0.59), while non-linear regression models resulted in heterocedasticity, non-normality and clustered data. NFIS could be a useful tool for estimating individual air pollution exposures in epidemiological studies on large populations, when emission-dispersion data are not available. The tradeoff between simplicity and accuracy needs to be considered.