Low Power Wide Area Networks (LPWAN) are very promising for a variety of IoT applications, but they face two major challenges: energy consumption of the wireless nodes and congestion of the networks due to the huge number of nodes involved. LPWAN transmission parameters can be optimised, e.g. using artificial intelligence algorithms, but the performance estimation made during simulations is often higher than what it is in reality. In this paper, we propose a new LoRa network simulator, J-LoRaNeS. Based on the Julia programming language, it allows fast prototyping and is therefore suited to the study of different adaptive LoRa mechanisms. We used our novel simulator to clearly show the benefits of using multi-arm bandits for adaptation, but we also show that the benefits reported in the literature are not attainable when realistic network conditions are simulated. CCS CONCEPTS • Computing methodologies → Simulation tools; Reinforcement learning.