Background: Urbanization and the increasing volume of municipal solid waste challenge traditional fixed-route collection strategies. Sensor-based smart waste systems offer a solution, but are costly to deploy at scale. This thesis explores a cost-effective alternative by combining machine learning (ML) for fill-level forecasting and evolutionary multi-objective optimization for waste collection routing, specifically targeting multi-compartment vehicles without relying on real-time sensors.Objectives: The primary objective is to develop a modular framework comprising(1) a sensor-free waste fill-level prediction module based on sequence-to-sequence neural networks, and (2) a multi-objective route optimization module using a customized Non-dominated Sorting Genetic Algorithm II (NSGA-II). The goal is to accurately forecast container fill levels and to generate efficient collection routes that minimize both travel distance and service time under realistic operational constraints.Methods: A synthetic fill-level dataset was generated to reflect realistic seasonality, holidays, and operational conditions. Transformer, LSTM, and BidirectionalLSTM models were trained for sequence-to-sequence forecasting. Route optimization employed a multi-objective NSGA-II algorithm, with a constraint programming pre-processor (CP-SAT) to assign valid collection days based on waste type, service frequency, and vehicle capacity constraints. Comparative benchmarking was conducted against Ant Colony Optimization (ACO) and a commercial PTV routing engine.Results: Transformer-based models consistently outperformed LSTM baselines in forecasting accuracy across synthetic scenarios. In routing, NSGA-II generated Pareto-optimal solutions balancing distance and service time, outperforming ACO and achieving competitive results compared to the commercial PTV solver. The combined framework demonstrated that predictive, sensor-free optimization is a viable pathway to smarter and greener waste collection systems.Conclusions: The findings confirm that sensor-free fill-level forecasting combined with evolutionary multi-objective optimization offers a scalable, cost-effective alternative for modern waste management. The modular approach allows for future integration with real-time systems and supports proactive planning strategies that reduce environmental and operational costs.