Speaker
Description
Abstract – Reinforcement Learning (RL) has demonstrated strong performance in production and manufacturing systems, where decision processes can be modeled as single-agent environments with welldefined state and action spaces. However, more complex real-world production processes inherently involve multiple interacting entities, making MultiAgent Reinforcement Learning (MARL) the natural tool of optimization. Even if similar to RL, MARL introduces new challenges, including non-stationarity, scalability issues, and increased training complexity due to agent interactions. This paper investigates an alternative approach that employs a single-agent RL framework within a multiagent environment. Instead of learning multiple coordinated policies, a single agent is trained to control multiple entities, with shared information. The study explores varying degrees of action and observation overlap between the controlled entities, analyzing how shared and independent components influence learning efficiency and policy quality. By comparing these configurations on the same test environment, the work provides insights into whether single-agent abstractions can effectively approximate multi-agent behavior. The results highlight the trade-offs between model simplicity and representational complexity, raising avareness for applying RL in complex production systems without the full overhead of MARL.