In order to estimate alpha-amylase, glucoamylase, acid proteinase, and acid carboxypeptidase activities in koji from the process variables and initial conditions of the koji making process, artificial neural network (ANN) models (ANN-10, -11, -15, and -21) were constructed with 10, 11, 15, and 21 input variables, respectively. These models could estimate the enzyme activities with high accuracy. Temperature and humidity orbits were then acquired by a genetic algorithm searching in the reverse direction using ANN-10, -11, -15, and -21 (GA-10, -11, -15, and -21). The orbits acquired by GA-15 and -11 were almost identical to the actual orbits, but those acquired by GA-21 and -10 were different. Enzyme activities acquired by GA-15 had 1.3% errors compared with the target values, while those acquired by GA-11 had 9.7% errors. GA-15 was, therefore, selected as the most suitable algorithm and was used to determine temperature and humidity orbits for target enzyme activities. Test koji making was then carried out according to the orbits acquired. As a result, the enzyme activities of the koji produced were almost the same as the target values.