Method for improving throughput of reinforcement learning system

• Main Authors: XIN YUSONG, LI KEQIU, HU YITAO, ZHAO ZHIXIN, DAI XIN'AN, ZHAO LAIPING
• Format: Patent
• Language: Chinese; English
• Published: 14.07.2023
• Subjects: CALCULATING; COMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS; COMPUTING; COUNTING; PHYSICS

The invention discloses a method for improving throughput of a reinforcement learning system, and the method comprises the following steps: starting an RL training task, deriving optimal global configuration by a coordinator according to task configuration and hardware information, and then starting an assembly line sampler, a quantizer and a trainer; the sampler carries out group-based parallel assembly line sampling and collects a certain number of tracks; the tracks are collected and distributed to a plurality of trainers by the message agent and are responsible for serializing and spreading the messages; the training device and the predictor use the received trajectory to train and evaluate the model, and send the updated model weight to a quantizer for weight quantization; and the quantizer sends the quantized model weight to the agent in each sampler through the message agent so as to carry out the next round of sampling and training. According to the method, the throughput of the reinforcement learning