We present a stochastic learning model that combines the essential elements of Hebbian and Rescorla-Wagner theories for operant conditioning. The model was used to predict the behavioral data of rats performing a vibrotactile yes/no detection task. Probabilistic nature of learning was implemented by trial-by-trial variability in the random distributions of associative strengths between the sensory and the response representations. By using measures derived from log-likelihoods (corrected Akaike and Bayesian information criteria), the proposed model and its subtypes were compared with each other, and with previous models in the literature, including reinforcement learning model with softmax rule and drift diffusion model. The main difference between these models was the level of stochasticity which was implemented as associative variation or response selection. The proposed model with subject-dependent variance coefficient (SVC) and with trial-dependent variance coefficient (TVC) resulted in better trial-by-trial fits to experimental data than the other tested models based on information criteria. Additionally, surrogate data were simulated with estimated parameters and the performance of the models were compared based on psychophysical measures (A': non-parametric sensitivity index, hits and false alarms on receiver operating characteristics). Especially the TVC model could produce psychophysical measures closer to those of the experimental data than the alternative models. The presented approach is novel for linking psychophysical response measures with learning in a yes/no detection task, and may be used in neural engineering applications.