Rescue Oriented Simulation: Handling and Navigating Fires (ROSHAN)
In today's world, which is increasingly plagued by wildfires, understanding and combating these natural disasters is of crucial importance.
This is where Roshan comes into play, the simulation tool for vegetation fires that is based on a combination of cellular automata and reinforcement learning. Roshan, standing for Rescue-Oriented Simulation, Handling and Navigating Fires, aims to simulate the dynamics of wildfire spread and develop effective fighting strategies. Utilizing the Korean LandCover database, Roshan generates realistic landscape models and provides an interactive interface that allows users to observe and control fires in real-time. A unique feature of Roshan is the use of an agent trained through reinforcement learning that is trained to identify and control fires in real-time.
A unique feature of Roshan is the use of an agent trained through reinforcement learning that is trained to identify and extinguish fires in real-time. A unique feature of Roshan is the use of an agent trained through reinforcement learning that is trained to identify and extinguish fires in real-time. The integration of maps is done directly via OpenStreetMap, accessible through a menu within Roshan. Users simply mark the desired area on the map, after which the data can be directly loaded into the simulation.
This data can also be saved and later reloaded to test different scenarios or situations. The simulation integrates 11 distinct terrain classifications, each reflecting unique landscape characteristics derived from the Korean LandCover database. These classifications are essential in defining the model's environmental parameters, closely aligning with the respective fuel types found in different regions. The heart of Roshan is a fire spread model named FireSpin. It is based on cellular automata. These automata simulate complex phenomena through simple rules and local interactions. Unlike traditional models, in Roshan, each cell generates virtual particles that move across a terrain and can ignite further cells. The model differentiates between two types of particles, radiation particles, which spread in a circular pattern around the cell, and convection particles, which are primarily dispersed by the wind. The Roshan model is designed to simulate complex phenomena through simple rules and local interactions. Unlike traditional models, in Roshan, each cell generates virtual particles that move across a terrain and can ignite further cells. The quantity and type of particles released by a cell are determined by its terrain class. These classes also regulate the ignition speed and the duration until a cell is completely burned out. As we have seen, terrains such as forests and grasslands are represented with unique characteristics. Forests, depicted in deep green, have a longer duration to fully burn and require more time to catch fire, reflecting their density. They spread fire mainly through the air, mirroring the real-world behavior of tree-canopy fires. Grasslands, shown in light 10, catch fire more quickly and burn rapidly, consistent with the behavior of dry, less dense vegetation. Though they spread fire less through the air, the heat from grassland fires can still ignite nearby areas. The heat from grassland fires can still ignite nearby areas.
At the core of Rochan's approach to wildfire management is the incorporation of a reinforcement learning agent represented by a drone.
This agent plays an important role in the system, offering a method to not only detect but also counteract the spread of fires within the simulation's environment. Using the principles of reinforcement learning, the agent continuously learns and adapts from its interactions with the simulated wildfires, aiming to operationally increase heat burn rates and be reducing the potential time taken to utilize the system performance performance.
These possibilities are important, both for game play described and for four-person play. We also want to point out that Roshan and its related developments are available as project on github. (translated with whisper)