EVOLUTIONARY TUNING OF HYPERPARAMETERS OF A HYBRID NAVIGATION ALGORITHM FOR ADAPTING A GROUP OF ROBOTS TO THE ENVIRONMENT TYPE

The efficiency of cooperative navigation algorithms for groups of mobile robots critically depends on their hyperparameters, the optimal values of which vary with the structure of the environment (e.g., office or polygon). Manually tuning these parameters for each new environment type is impractical. This paper proposes a method for automatically adapting the behavior of a robot group using evolutionary tuning. Based on the GARCA hybrid architecture, which considers neighbors' goals, a compact vector of six interpretable hyperparameters is formed. A genetic algorithm with real-valued encoding is used for their optimization. The results of a numerical experiment confirm that evolutionary tuning can automatically discover different behavioral strategies: an "aggressive" one for structured office environments and a "cooperative" one for environments with chaotic obstacles. The adapted GARCA+GA algorithm significantly outperforms methods with fixed parameters in terms of mission success rate and completion time.