Clarifying the distribution of radioactive sources within nuclear facilities is essential for ensuring worker safety during decommissioning and emergency responses. However, air dose rate measurements are often limited in complex and highly contaminated areas. To address this, we propose an optimized machine learning-based approach using the Least Absolute Shrinkage and Selection Operator (LASSO) for reconstructing radioactive source distributions. While LASSO performs well in simple room models, its accuracy diminishes in more complex environments due to obstacles and shielding effects. To overcome this, we developed an optimized LASSO scheme that normalizes radioactive contributions from sources, mitigating the impact of shielding. Numerical simulations demonstrate that the optimized approach significantly improves reconstruction accuracy compared to the non-optimized version. Experimental validation in a complex room further confirms the effectiveness of the method. This optimized LASSO scheme shows promise for future applications in monitoring and decommissioning nuclear facilities, providing high accuracy in both operational and damaged environments.