This study introduces the integration of three artificial intelligence techniques to establish a communication system in which agents learn from their environment using a deep learning algorithm and then share the acquired knowledge. By communicating, the agents can utilize the insights gained by others in the system, reducing the number of actions that can lead the agent to a non-ideal or prohibited situation. The approach incorporated a pair-based multi-agent architecture with clearly established roles and involved some modifications to the Deep Reinforcement Learning algorithm used which was Deep Q-learning. One modification involves adding a flag in the observations within the experience replay. This flag enables the agent to identify a relevant state making it possible to recalculate the expected values during the network training. Additionally, limiting the use of epsilon-greedy prevented the agents from further exploring states where errors had previously been made. These modifications proved their effectiveness since the number of episodes where mistakes were made was reduced by around 92.