— in these days, due to computational complexity and time consuming of classical methods, researches tend to using artificial intelligence approaches such as reinforcement learning method, which are the most convenient ways of dealing with sophisticated problems, specially knowing the fact that they do require no prior knowledge about the problem. In this research, a navigation algorithm based on reinforcement learning is proposed. In order to use reinforcement learning techniques, the main steps are to design the required states capable of representing the ongoing world realistically and a reward function based on which learning can happen. When the aim is to put the reinforcement learning techniques into work in dynamic environments, the designing of states and reward function should be done very delicately. Due to complexity of dynamic environments, finding a mapping function from environment status to states is a very challenging task. In addition, dynamics nature of the environment can also affect the reward signal as well and it will cause the variation in the amount of reward signal which a learning agent receives at each moment. As reinforcement learning methods are all based on reward functions, the quality of learning can be very dependent on the received reward signal during the time which the learning supposed to be happening. In such unsupervised learning method, the data is going to be provided by the environment, and not by a supervisor, and due to the randomness of such provided data, the learning agent will be faced with inappropriate convergence of algorithm, even if it was not diverged as there’s no guarantee for that to happen. In the current study, besides bringing the proof of proposed method convergence and describing and analyzing the manner in which an appropriate policy can be formed through the learning process, a heuristic function will be introduced to speed up the whole learning process. As is well-known the popularity of reinforcement learning methods is all because of their eye-catching performances, especially in unknown environments; so no prior knowledge is needed to be available on the agent’s status and its capability, and also the environment situation. In order to define a heuristic function which any prior information will be stay avoided, the definition of heuristic function and the guidance of agent in the environment is solely based on the acquired data from the environment. Such definition causes the heuristic function to be affected by the provided data. In every task, there are some certain conditions which must be held under any circumstances or otherwise it will be counted as system failure. The aim of adjusting the weights of a multi-objective reward function is exactly to control the agent’s behavior and furthermore, to prevent any system failure; so the consequences due to data dependency of huristic function in use can be tackled in the same way. At the end, the simulation results shows that the proposed approach has higher performance than any other previous researches were conducted in this literature, and more importantly, the performance of learning agent keeps improving through the learning process and there’s no casual reduction in performance, so the convergence of algorithm is guaranteed. As it was expected in analyzing the proposed method, an appropriate policy will be shaped in a very short time through learning process, no matter how the training data are being represented. Hence, the user can use the proposed approach free of any further concerns. Keywords: Navigation, dynamic environment, obstacle avoidance, reinforcement learning