Studies on psychiatric disorders in abnormal psychology have shown several cognitive and affective features of psychiatric patients such as attentional bias, deficits in working memory, deviations in the parameters of reinforcement learning such as learning rate, inverse temperature, and discounting rate. However, causal relationships between such cognitive and affective features and symptoms of the disorders are unclear. The hypothesis of this study is that the dysregulation of homeostasis and allostasis via mechanisms of the predictive coding of interoception may be a critical mediator of the link between cognitive and affective features and psychiatric disorders. In this paper, a computational model combining the predictive coding of interoception and reinforcement learning is proposed to provide suggestions for the hypothesis. Simulations using the model suggested that (1) a reduced learning rate and inverse temperature, which are observed in depressed patients, can lead to unstable decision-making and maintained higher levels of reward predictive errors and (2) can consequently result in dull physiological reactivity and chronically higher levels of autonomic responses. These results provide a perspective that can integrate cognitive and affective features, physiological states, and symptoms of psychiatric disorders.