Poultry processing plants process broilers into products ordered by customers such as retailers and wholesalers. In this thesis, we focus on the batching process of such plants, where products are processed into packages of products (batches) that one can find in a supermarket. Batchers process arriving products into one of multiple available bins, producing batches for retailers. Alternatively, products can be passed down to a batcher producing batches for wholesalers. Retailers demand batches with a given minimum target weight and throughput, while paying a fixed price per batch. Additional weight per batch is therefore not paid for, and also known as giveaway. Instead, wholesalers pay for all weight in a batch, but less per gram. We consider two variations of this on-line decision problem. In the first case, we only know the weight of the next arriving item. We model this setting as a Markov decision process (MDP), and develop heuristics for realistic settings as the MDP scales poorly with the problem size. The second case is considered semi-online, where we know the weight of multiple arriving items instead. We develop a hybrid genetic algorithm (HGA) to utilize the extra information on product weights. The available calculation time to make a decision is constrained, due to the high capacity of this type of batcher. In a numerical study we show that the HGA significantly outperforms current practice in terms of giveaway and achieving the target throughput. On the planning level it must be taken into account that broiler weights vary over time, as the relation between target weight and product weight distribution strongly influences giveaway performance. Moreover, multiple jobs need to be produced on one or more batchers, each job with its own number of batches, target weight, and due date. We model this setting by incorporating the giveaway performance of the batchers, resulting in a scheduling problem where jobs have time-dependent processing times. The problem is tackled using both a mixed-integer linear program (MILP) and a HGA, which incorporate the tardiness and giveaway minimization objectives. The MILP is unable to solve large problem instances, but the HGA is shown to outperform current practice, also for large problem instances.