Automation Efficiency in Cyber-Physical Systems: Interaction Between Technical Performance and Algorithmic Intelligence
The performance of a cyber-physical system cannot be judged reliably from a single engineering variable. In the investigated micro-CPS, response time, communication latency and throughput describe the technical layer, whereas prediction accuracy captures the contribution of the learning component. The present paper examines how these two dimensions interact and how their combined effect is reflected in an overall efficiency index. A quasi-experimental comparison was carried out for a logistics and warehouse control configuration equipped with connected sensing, real-time data exchange and machine-learning support, and for a functionally similar conventional automated system. The empirical analysis was based on six variables, which were first standardized using z-scores and then assigned to four domains: technical efficiency, algorithmic intelligence, resource and energy efficiency, and economic efficiency. A clear difference emerged between the two system architectures. The CPS produced a mean composite index of 0.954, whereas the conventional automated system had a mean value of −0.954. The regression results further showed that technical efficiency and algorithmic intelligence contributed most strongly to the overall index. Of particular interest was their interaction. The interaction between technical efficiency and algorithmic intelligence was statistically significant (B = 0.154, β = 0.267, p = 0.004). This interaction included to the model increased R² from 0.93 to 0.95. The finding indicates that the contribution of algorithmic intelligence to overall system performance depends, at least partly, on the level of technical efficiency achieved by the system. Rather, their contribution becomes more pronounced when the control and communication infrastructure already performs reliably. The sensitivity analysis led to the same overall ordering of the two architectures under alternative weighting schemes. Within the scope of the available dataset, this provides additional evidence that the observed difference is not simply an artefact of the selected weights.