Deep reinforcement learning robust training method and device based on neuron coverage rate

Abstract

The invention discloses a deep reinforcement learning robust training method and device based on neuron coverage rate, and the method comprises the following steps: (1) building an intelligent driving environment, collecting state data from the intelligent driving environment, and training a deep reinforcement learning model until a set return value is reached; (2) performing operating in an environment by using the trained deep reinforcement learning model, and extracting multiple rounds of state action pairs; (3) constructing a predictor used for predicting a future moment state action pair sequence according to the historical state action pair sequence and a classifier used for carrying out quality classification on the state action pairs, and training the predictor and the classifier by utilizing the extracted state action pairs; and (4) according to the defined adversarial sampling strategy, performing quality sampling on the state action pair according to the state action pair, and performing retraining of the deep reinforcement learning model so as to improve the robustness of the deep reinforcement learning model.

Description

technical field [0001] The invention relates to the field of artificial intelligence, in particular to a robust training method and device for deep reinforcement learning based on neuron coverage. Background technique [0002] With the rapid development of artificial intelligence, deep reinforcement learning algorithms have become one of the most concerned algorithms in this field. Deep reinforcement learning combines the perception ability of deep learning with the decision-making ability of reinforcement learning, and can perform end-to-end control directly according to the input information, and solve the sequential decision-making problem in high-dimensional state space. Due to its excellent performance, deep reinforcement learning algorithms are widely used in autonomous driving, automatic translation, dialogue systems, and video detection. However, the neural network black box lacks interpretability, and it is difficult to guarantee the security. Therefore, it is very...

AI Technical Summary

Problems solved by technology

First of all, most of the training methods generate a large number of repetitive operations. When the agent first learns the task, it is easy to cause failure. With continuous learning, the frequency of failure will decrease.

The agent will continue to encounter the solutions it has mastered. The performance improvement of the agent at this stage is very small. Such training will lead to unreasonable data starvation

Images