Adversarial attack and defense in reinforcement learning-from AI security view

Cybersecurity

Table 1 The comparison of the most representation reinforcement learning algorithms

RL algorithm	Approach type	Learning type	Application scenarios
Q-Learning (Watkins and Dayan 1992)	Value-based	Shallow Learning	Motion Control, Control System,
			and Robot Application et al.
DQN (Mnih et al. 2013)	Value-based	Deep Learning	Motion Control, Neutralization Reaction
			Control, and Robot Path Planning et al.
VIN (Tamar et al. 2016)	Value-based	Deep Learning	Path Planning, and Motion Control et al.
A3C (Mnih et al. 2016)	Combined	Deep Learning	motion Control, Game Playing, self-driving,
			and Path Planning et al.
TRPO (Schulman et al. 2015)	Policy-based	Deep Learning	Motion Control, and Game Playing et al.
UNREAL (Jaderberg et al. 2016)	Combined	Deep Learning	Motion Control, and Game Playing et al.

Approach Type contains two categories, namely Policy-based, and Value-based. Meanwhile, learning Type also contains two categories, namely Shallow Learning, and Deep Learning