The mechanisms of decision-making are generally thought to be under the control of a set of cortico-subcortical loops involving basal ganglia and thalamic pathways. These structures include several parallel functional loops connecting back to distinct areas of cortex, processing motor, cognitive and limbic modalities of decision making. Due to convergence and divergence within the network, these loops cannot be completely segregated. We used these circuit properties to develop a computational model at a spiking neuron level of description. The model was implemented using leaky integrate-and-fire neuronal models connected by voltage-jump synapses and its architecture relied on commonly accepted data regarding the complex functional connectivity description between basal ganglia, cortex and thalamus. This model was applied to a decision making task which was initially studied in primates. In this task, the animals were trained to associate reward probabilities to different cues they had to select in order to maximize their gain. Combining behavioral and electrophysiological experimental data from this study and detailed circuit description, we developed a basal ganglia model in which we used two parallel loops, each of which performed decision making based on interactions between positive and negative feedback pathways. The loops communicate via partially convergent and divergent connections in specific basal ganglia sub-regions. This neuronal network model was then trained to perform the same decision making task as in primates. This training resulted from the closed-loop interaction between the neural circuitry and its sensory-motor interface. The learning was implemented as a cortico-striatal synaptic weight variation modulated by phasic dopamine release following the presence or absence of reward delivery. Thanks to this simple bottom-up approach the model was able to learn to select optimum reward cues in a similar manner as the monkey. Moreover, this model allows us (i) to avoid the arbitrary choice of a pre-existing machine-learning derivative model, (ii) to investigate lesional and pharmacological effects on learning and decision making and (iii) also provides the possibility to test for further hypotheses regarding cell-scale mechanisms effect on the whole model capacities.