Abstract:
This thesis proposes the framework to find the optimal selection of heterogeneous wireless network. Reinforcement learning (RL) model has been used to find the best strategy to maximise the reward function expressed in terms of call blocking and call dropping probabilities. The reward-evaluation model is based on the well-established macroscopic cell transmission model (CTM). In this regard, numerical results have been given on a simple 1-dimensional road network settings with choices of moving users in making their connections via micro and macro cells in both direct and ad hoc modes. With this approach, traffic sources have been modelled as the deterministic fluid flow of moving users travelling on their vehicles along a road. In contrast to the approaches with microscopic user mobility models often used in the past, the proposed framework has the advantage of computational efficiency and can be integrated well with RL in the herein developed optimisation framework. The proposed framework has been considered RL-sensitivity in order to find the sensitivity of RL-parameters and evaluated in different scenarios which are the changes of bandwidth, stochastic incoming demands and unpredictable network problems. The results show that the RL algorithm can lead to the optimal solutions in all the tested scenarios with no large computational complexity as the other algorithms which have the efficiency nearly equal to RL. Furthermore, the RL algorithm can improve the network selection automatically even if the topology of the system has been changed immediately.