In the Internet of Things (IoT) context such as Low Power Wide Area Network (LPWAN), it is essential to reduce the packet losses, e.g., to save energy. Decentralized artificial intelligence techniques have been proposed to combat radio collisions, but the approach here is extended to deal additionally with the channel propagation effects. In this article, a Quality of Channel Allocation (QoC-A) learning technique based on bandit algorithms is proposed in order to choose the transmission channel. This aims to reduce the effect of the propagation impairments between the radio channels while using the Effective Signal Power (ESP) as a quality metric. In addition, a Discounted QoC-A (DQoC-A) algorithm is proposed to adapt rapidly to any abrupt change in the channelsand#x2019; conditions. An experimental campaign on a real IoT device is carried out to demonstrate the low complexity and efficiency of these proposed decentralized algorithms. In the given results, QoC-A outperforms the classical UCB policy with a more accelerated learning process. On the other hand, the feasibility of using the DQoC-A in non-stationary scenarios is illustrated by its rapid convergence when abrupt changes in the channelsand#x2019; conditions occur. At the end of the process, these proposed learning techniques give 4.1 and 2.4 times fewer packet losses than the traditional ones with a random channel assignment scheme, in the stationary and non-stationary scenarios, respectively.