Reinforcement learning: a control approach for reducing component damage in mobile machines

This paper presents an active component damage reducing control approach for driving manoeuvres of a wheel loader. For this purpose, the front and rear axle loads will be manipulated by force pulses induced into the machine chassis via the lifting cylinders of the function drive. The associated control approach is based on the principles of Reinforcement Learning. The essential advantage of such methods against linear control approaches is that no descriptive system properties are required, but the algorithm automatically determines the system behaviour. Due to the high number of necessary training runs, the algorithm is designed and taught using a validated wheel loader simulation model. After over 850 training runs, an optimal strategy for damping the axle loads could not yet be determined. In spite of the unprecedented convergence, initial improvements of the damage values have already been achieved on tracks that deviate from the training track. Some of these results show a 4.9 % lower component damage compared to a machine setting with no damping system. The results and limits of this strategy are discussed due to a comparison with other scientific active vibration damping approaches. ... mehrCurrently, a linear control method (P-PI-controller) has a higher damage reduction potential, but it is expected that further training runs and another learning algorithm could make the reinforcement learning approach even more effective. Coupling the linear control method with the selflearning approach shows the highest potential for the axle damage reduction.