Abstarct: Reinforcement learning is an interesting area of machine learning, which aims at improving the agent behavior baxsed on Reinforcement signals received from the environment. However, in many real applications, the reward is given to the agent with much delay. As a result, the agent needs to spend much time to achieve optimal behavior. Different methods such as reward shaping have been proposed to overcome this problem. But none could have noticeable effect to increase the learning speed, especially in large and real environments. Another problem is that as long as the agent doen not reach to an acceptable level of learning, all of its movements are random. Moreover, further complicating the environment would lead to an increase in the number of explorations and decision parameters. These issues make exploration time consuming, costly and sometimes very dangerous. An interesting solution for researchers to address these issues is qualitative learning. In this thesis, a general frxamework for qualitative learning, along with its properties and components is presented. This frxamework is designed baxsed on the qualitative learning and reward shaping to take the benefits from both worlds and provides as much possible convergence rate as possible. The proposed frxamework is adjustable and adaptable to different algorithms either discrete or continuous, and navigation and non-navigation environments. Then, a prototype is instantiated from the proposed frxamework and further evaluated on some benchmark environments. The results of the experiments demonstrate the effectiveness of the proposed frxamework to achieve the optimal policy and accelerating the learning process.