This thesis addresses the high amount of uncertainty present when deciding on an optimal strategy for capacity expansions in the biopharmaceutical industry. Uncertainties pertaining to demand, yield, and clinical trial success are considered and evaluated through three different solution frameworks: deterministic Mixed Integer Linear Programming (MILP), Stochastic Programming (SP) with Recourse, and a Markov Decision Process (MDP) trained with Deep Reinforcement Learning (DRL). The results show, that the SP model quickly becomes infeasible due to the high dimensionality of the problem. The Reinforcement Learning (RL) based model outperforms the MILP model, but the quality of it’s solutions also deteriorates drastically when complexity gets too high. Therefore, while the potential of this solution approach that allows for stochasticity in highly combinatorial observation spaces is shown, an improvement of the policy network’s architecture is necessary for the RL model to be a viable option for industrial use.     «

This thesis addresses the high amount of uncertainty present when deciding on an optimal strategy for capacity expansions in the biopharmaceutical industry. Uncertainties pertaining to demand, yield, and clinical trial success are considered and evaluated through three different solution frameworks: deterministic Mixed Integer Linear Programming (MILP), Stochastic Programming (SP) with Recourse, and a Markov Decision Process (MDP) trained with Deep Reinforcement Learning (DRL). The results sho...     »