Deep reinforcement learning model robustness enhancement method based on information bottleneck

Abstract

The invention discloses a deep reinforcement learning model robustness enhancement method based on information bottleneck. According to the deep reinforcement learning model robustness enhancement method based on the information bottleneck, state information in deep reinforcement learning is limited by setting the information bottleneck, the state information in a transfer tuple is encoded through an encoder, firstly, the state observed in the environment is encoded, encoding the data, inputting the encoded data into a strategy network, interacting with the environment according to the action of the strategy network to obtain the state of the next round, encoding the state, and continuously interacting with the environment to realize the training of the strategy network. According to the deep reinforcement learning model robustness enhancement method based on the information bottleneck disclosed by the invention, a strategy obtained by training still has good performance on an original task, and the influence of adversarial attacks can be resisted; a proportionality coefficient in a regular term is set by adopting an annealing thought, so that a stable training process is achieved, and a strategy obtained by training still has excellent performance in a normal task.

Description

technical field [0001] The invention relates to the field of enhanced robustness of deep reinforcement learning; in particular, it relates to a method for enhancing the robustness of a deep reinforcement learning model based on information bottlenecks. Background technique [0002] With the rapid development of artificial intelligence, deep reinforcement learning algorithms that combine the perception ability of deep learning and the decision-making ability of reinforcement learning are widely used in automatic driving, automatic translation, dialogue systems and video detection. [0003] However, deep reinforcement learning combined with deep learning is vulnerable to adversarial attacks, which add some noises that are imperceptible to the human eye to the original samples, which do not affect human recognition, but enable the trained policy to make An action that is extremely unfavorable to the outcome, leading to the failure of the entire decision-making process. [0004...

AI Technical Summary

Problems solved by technology

[0003] However, deep reinforcement learning combined with deep learning is vulnerable to adversarial attacks, which add some noises that are imperceptible to the human eye to the original samples, which do not affect human recognition, but enable the trained policy to make An action that is extremely unfavorable to the outcome, leading to the failure of the entire decision-making process

Images