Analýza a porovnání algoritmů pro hledání tras nad prostorovými daty

This thesis addresses automated routing of high voltage transmission lines (VVN/ZVN) in the Czech landscape and proposes, implements, and experimentally evaluates a methodology that compares a deterministic and an adaptive routing approach over a unified geospatial data base. The cost map is constructed from 32 thematic layers of the ZABAGED database, a digital terrain model, and a slope raster, complemented by mechanisms for inward gradients of protected areas, buffer zones around built-up land, and slope penalization. The first routing approach is Dijkstra's algorithm operating on an extended state space that incorporates the incoming direction and the number of long-jump edges used, providing a mathematically optimal reference path. The second approach is deep reinforcement learning in the form of the MaskablePPO algorithm; the routing task is formulated as a Markov Decision Process, and the agent is trained using the Stable-Baselines3 library. Sensitivity of the resulting path to cost-map construction is examined in four isolated experiments (protected-area gradient, built-up buffer shape, slope penalty, land-cover weights). The results show that path geometry is primarily driven by land-cover weights and by the location of impassable polygons, while slope penalty and built-up buffers act locally. Dijkstra's algorithm yields deterministically optimal and geometrically smoother routes, whereas the PPO agent produces longer and costlier routes but offers stochastic solution variability, fast inference independent of the area size, and demonstrable generalization to unseen landscape configurations without retraining.