DescriptionThermal management is one of the very important technique in modern mobile SoC designs. Applying reinforcement learning to thermal management has a high potential in terms of performance improvement, but has a disadvantage that it is very difficult to train in real environment. This research proposes a virtual environment to train a reinforcement learning. Through this environment, reinforcement learning training can be freely performed without any restrictions that may occur in the real world, and optimal reinforcement learning techniques can be derived. Experimental results showed that the thermal management based on reinforcement learning using the proposed method can manage heat more effectively while showing 3% higher performance than the existing thermal management method.