A PCA-Based Model to Predict Adversarial Examples on Q-Learning of Path Finding

摘要

Reinforcement learning is a core learning approach of machine learning in artificial intelligence, which has been widely used to realize the module of automatic path finding for robots. However, recent researches have shown that the designed adversarial examples can make an algorithm-level attack on reinforcement learning in the Atari game. Obviously, such attack may be also brought into the automatic path finding based on reinforcement learning. In this paper, we focus on the adversarial example-based attack on a representative reinforcement learning named Q-learning in automatic path finding. We propose a probabilistic output model based on the influence factors and the corresponding weights to predict the adversarial examples. Through calculation on five factors including the energy point gravitation, the key point gravitation, the path gravitation, the included angle and the placid point, a natural linear model is constructed to fit these factors with the weight parameters computation based on the principal component analysis(PCA). Through massive experiments, we successfully find the adversarial examples for the first time on Q-learning in path finding and our model can make a satisfactory prediction. Under a guaranteed recall, the precision of the proposed model can reach to 70% with the proper parameter setting.