Integrated high frequency RF circuit design using deep reinforcement learning via proximity policy optimization method

The automatic design of analog circuits is a challenging task due to the high complexity of the design, which is caused by the search space and the sometimes conflicting parameters. In the article, a trial-and-error-based approach that combines reinforcement learning and deep neural networks is used to determine the values of circuit elements have been used. In methods based on reinforcement learning, the agent tries to act like an expert designer and maybe better than that by trial and error and using the information he gets from the environment. In this article, one of the latest methods of deep reinforcement learning called approximate policy optimization (PPO) is used. To show the efficiency of the above method, a cascaded LNA circuit is considered. And the voltages are determined by the learning agent to optimize the circuit design requirements such as gain, noise figure and power consumption. To train the learning agent in the reward function, two categories of adverbs have been included in such a way that the main goal is to optimize the gain and noise figure and the secondary goal is to focus on other requirements such as power consumption. The environment which is the amplifier circuit is simulated in the Hspice software in ۰.۱۸ micrometer technology from TSMC company at the frequency of ۵.۷ GHz and the learning agent is also defined in the MATLAB environment which has been able to design the values of the circuit elements by interacting with the environment.

کلیدواژه ها:

integrated circuit design ، RF ، deep reinforcement learning ، PPO ، LNA