The article aims to analyze three different functionals used as error norms in reconstructing the thermal conductivity coefficient κ under fourth-kind boundary conditions. The thermal conductivity coefficient κ is significant in modeling heat conduction at the interface of the casting and the mold. The study utilized two optimization algorithms (ABC – Artificial Bee Colony and ACO – Ant Colony Optimization), with the primary objective being to assess how individual functionals respond to input data disturbances and how they affect the quality and stability of the parameter selection process. The studies were conducted for three levels of disturbances: 0%, 1%, and 2% of the input data and various population sizes: 5, 10, 15, and 20, with a constant number of 6 iterations. In each case, the average value of the functional and the compliance of the determined parameter κ with the reference value were analyzed. The efficiency of the algorithms was compared in terms of result stability and data error resistance. The L1 functional demonstrated high regularity and stability regardless of computational conditions. ABC and ACO generated consistent solutions, although ACO stood out as having more excellent repeatability. The L2 functional, more sensitive to disturbances, showed significant differences between the algorithms: ABC reacted strongly to disturbances, leading to a scatter of results, while ACO maintained high mapping quality. The most important differences were observed in the case of the L∞ functional, which exhibited the highest susceptibility to data disturbances. In this case, only ACO could maintain the stability and precision of the solutions, indicating its higher resilience in the context of mapping the parameter κ. The analysis shows that the nature of the functional is crucial for the entire optimization process. It is not the algorithm, but the properties of the functional that determine the difficulty of the inverse problem and the scale of the final error. The selection of the functional should be treated as a strategic decision affecting the convergence, stability, and reliability of the obtained solutions. These conclusions are crucial in the context of engineering applications, where the precision and robustness of the computational method directly translate into the quality and safety of the technological process.