BLACK-BOX OPTIMIZATION METHODS FOR REAL-TIME PARAMETER TUNING IN ROBOTICS

Kurzfassung

Hyperparameter tuning is essential for the optimal functioning of robotic systems. Currently, this task is often performed by experts using a combination of heuristics and random search, a time-consuming, hard-to-replicate process. Finding appropriate hyperparameters is an even more pronounced problem on robots with a high number of degrees of freedom, such as humanoids. This work proposes to automate tuning procedures using black box optimization, in particular the CMA-ES algorithm. Adaptations are proposed to the standard form of CMAES to increase performance on tasks relevantly similar to the use case. A Gaussian process surrogate assisted version of CMA-ES with an epistemic-uncertainty-based step-size modulation mechanism is proposed. The real world use case consists of tuning a humanoid robot to perform a running task with minimum energy expenditure. Algorithms are first tested on canonical black box optimization benchmarking functions and subsequently in simulation. The developed algorithms resulted in the finding of hyperparameters yielding up to a 10% lower cost than that obtained using hyperparameters found by hand tuning in the running task. These results show that black box optimization through CMA-ES variants is a viable research avenue for hyperparameter tuning in 30-dimensional spaces.