A computer-implemented method of learning a policy for controlling a computercontrolled system, in particular a robot, by reinforcement learning, comprising the steps of: Observing a current state (s) of an environment of the computercontrolled system. Interacting (S12) with the environment by carrying out the steps of: Determining an action (a) by the policy, which is an actor network, depending on the current state (s), Executing the action (a) by the computercontrolled system, receiving a reward (r) and the next state (s'), and storing the interaction with the environment as an experience in a replay buffer. Sampling (13) a batch of experiences from the replay buffer. Computing (S15) Q values by at least two critic networks for the experiences in said batch. Determining a standard derivation of the Q values for the experiences in said batch for each of the critic networks. Updating (S17) critic parameters depending on the Bellman equation and adding the standard derivations as regularization.
DEVICE AND METHOD TO IMPROVE REINFORCEMENT LEARNING
VORRICHTUNG UND VERFAHREN ZUR VERBESSERUNG DES VERSTÄRKUNGSLERNENS
DISPOSITIF ET PROCÉDÉ POUR AMÉLIORER L'APPRENTISSAGE PAR RENFORCEMENT
2025-04-09
Patent
Electronic Resource
English
IPC: | G05B Steuer- oder Regelsysteme allgemein , CONTROL OR REGULATING SYSTEMS IN GENERAL / B60W CONJOINT CONTROL OF VEHICLE SUB-UNITS OF DIFFERENT TYPE OR DIFFERENT FUNCTION , Gemeinsame Steuerung oder Regelung von Fahrzeug-Unteraggregaten verschiedenen Typs oder verschiedener Funktion / G06N COMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS , Rechnersysteme, basierend auf spezifischen Rechenmodellen |