Grid geometry is a foundational but rarely examined design choice in Reinforcement Learning (RL) environments for spatial exploration. We introduce HexaWorld and SquareWorld, two structurally parallel RL environments differing exclusively in their underlying discretization, and present a controlled comparison of hexagonal versus square grids under partial observability and finite energy budgets. Both environments are instantiated over three Mediterranean regions using Copernicus Marine Service and EMODnet data, with composite stress indices operationalized as reward signals. Across 135 matched configurations trained with MaskablePPO, HexaWorld outperforms SquareWorld in 79% of cases, with the advantage scaling with energy budget and habitat depth. A per-biozone analysis reveals a depth-dependent dissociation: SquareWorld achieves higher exploration efficiency in shallow zones, while HexaWorld translates coverage into reward more effectively at circalittoral and bathyal depths, suggesting hexagonal discretization as the stronger choice for deep benthic stress monitoring.
HexaWorld: The Role of Grid Geometries in Constrained RL Exploration / Lombardi, G., Bianchi, L.A., Schneider, R.. - 16978:(2027), pp. 115-127. (12th IAPR TC3 Workshop on Artificial Neural Networks in Pattern Recognition, ANNPR 2025 ita 2026) [10.1007/978-3-032-39028-8_10].
HexaWorld: The Role of Grid Geometries in Constrained RL Exploration
Giulia Lombardi
;Luigi Amedeo Bianchi;
2027-01-01
Abstract
Grid geometry is a foundational but rarely examined design choice in Reinforcement Learning (RL) environments for spatial exploration. We introduce HexaWorld and SquareWorld, two structurally parallel RL environments differing exclusively in their underlying discretization, and present a controlled comparison of hexagonal versus square grids under partial observability and finite energy budgets. Both environments are instantiated over three Mediterranean regions using Copernicus Marine Service and EMODnet data, with composite stress indices operationalized as reward signals. Across 135 matched configurations trained with MaskablePPO, HexaWorld outperforms SquareWorld in 79% of cases, with the advantage scaling with energy budget and habitat depth. A per-biozone analysis reveals a depth-dependent dissociation: SquareWorld achieves higher exploration efficiency in shallow zones, while HexaWorld translates coverage into reward more effectively at circalittoral and bathyal depths, suggesting hexagonal discretization as the stronger choice for deep benthic stress monitoring.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione



