In the LTE access networks, the Radio Resource Management (RRM) is one of the most important modules which is responsible for handling the overall management of radio resources. The packet scheduler is a particular sub-module which assigns the existing radio resources to each user in order to deliver the requested services in the most efficient manner. Data packets are scheduled dynamically at every Transmission Time Interval (TTI), a time window used to take the user’s requests and to respond them accordingly. The scheduling procedure is conducted by using scheduling rules which select different users to be scheduled at each TTI based on some priority metrics. Various scheduling rules exist and they behave differently by balancing the scheduler performance in the direction imposed by one of the following objectives: increasing the system throughput, maintaining the user fairness, respecting the Guaranteed Bit Rate (GBR), Head of Line (HoL) packet delay, packet loss rate and queue stability requirements. Most of the static scheduling rules follow the sequential multi-objective optimization in the sense that when the first targeted objective is satisfied, then other objectives can be prioritized. When the targeted scheduling objective(s) can be satisfied at each TTI, the LTE scheduler is considered to be optimal or feasible. So, the scheduling performance depends on the exploited rule being focused on particular objectives. This study aims to increase the percentage of feasible TTIs for a given downlink transmission by applying a mixture of scheduling rules instead of using one discipline adopted across the entire scheduling session. Two types of optimization problems are proposed in this sense: Dynamic Scheduling Rule based Sequential Multi-Objective Optimization (DSR-SMOO) when the applied scheduling rules address the same objective and Dynamic Scheduling Rule based Concurrent Multi-Objective Optimization (DSR-CMOO) if the pool of rules addresses different scheduling objectives. The best way of solving such complex optimization problems is to adapt and to refine scheduling policies which are able to call different rules at each TTI based on the best matching scheduler conditions (states). The idea is to develop a set of non-linear functions which maps the scheduler state at each TTI in optimal distribution probabilities of selecting the best scheduling rule. Due to the multi-dimensional and continuous characteristics of the scheduler state space, the scheduling functions should be approximated. Moreover, the function approximations are learned through the interaction with the RRM environment. The Reinforcement Learning (RL) algorithms are used in this sense in order to evaluate and to refine the scheduling policies for the considered DSR-SMOO/CMOO optimization problems. The neural networks are used to train the non-linear mapping functions based on the interaction among the intelligent controller, the LTE packet scheduler and the RRM environment. In order to enhance the convergence in the feasible state and to reduce the scheduler state space dimension, meta-heuristic approaches are used for the channel statement aggregation. Simulation results show that the proposed aggregation scheme is able to outperform other heuristic methods. When the aggregation scheme of the channel statements is exploited, the proposed DSR-SMOO/CMOO problems focusing on different objectives which are solved by using various RL approaches are able to: increase the mean percentage of feasible TTIs, minimize the number of TTIs when the RL approaches punish the actions taken TTI-by-TTI, and minimize the variation of the performance indicators when different simulations are launched in parallel. This way, the obtained scheduling policies being focused on the multi-objective criteria are sustainable. Keywords: LTE, packet scheduling, scheduling rules, multi-objective optimization, reinforcement learning, channel, aggregation, scheduling policies, sustainable.