The rapid scaling of large language models (LLMs) has significantly increased GPU memory pressure, which is further aggravated by training optimization techniques such as virtual pipeline and recomputation that disrupt tensor lifespans and introduce considerable memory fragmentation. Such fragmentation stems from the use of online GPU memory allocators in popular deep learning frameworks like PyTorch, which disregard tensor lifespans. As a result, this inefficiency can waste as much as 43% of memory and trigger out-of-memory errors, undermining the effectiveness of optimization methods.

Poor visibility by sea fog is a navigational hazard for maritime transport. Due to the lack of observation stations on the sea, monitoring large-scale sea fog and assessing navigation risk in foggy season are challenges. Thus, the paper introduced MODIS data for monitoring sea fog and proposed an approach of assessing navigation risk along fairway during foggy season. Sea fog is firstly identified from MODIS images using supervised classifier and then navigation risk along fairway is quantitatively evaluated by means of spatial analysis. Using GIS based Multi-criteria Decision Analysis (MCDA) model, a risk index layer is respectively constructed for various marine environmental factors, which include sea fog, wind, wave and etc. and these risk index layers eventually are integrated into navigation risk assessment layer through the order weighted average (OWA) operator.

In this paper, we proposed a two-phase many-objective evolutionary algorithm to tackle many objective optimization problems. In the first phase, the algorithm focuses on achieving good convergence towards the boundary Pareto optimal solutions. In the second phase, it maintains a good balance between convergence and diversity by using a set of widely spread reference lines. In addition, a penalty based adjustment for reference line has been adopted to handle many objective optimization problems with incomplete PFs. The performance of our proposed algorithm is validated and compared with four state-of-the-art many objective evolutionary algorithms on DTLZ problems. The results show that our proposed algorithm is very competitive with other compared algorithms.