Table 2 Comparison of recent DRL-based collision avoidance Methods.

From: Deep reinforcement learning model for Multi-Ship collision avoidance decision making design implementation and performance analysis

Research

Learning architecture

State space features

Action Space

Key contribution

Limitation

Zhao et al9.

Single-agent DQN

Relative positions, COLREG situations

Discrete heading changes

First application to restricted waterways

Limited to 3 vessels, no speed actions

Li et al10.

Multi-agent DRL

Local observations, partial information

Heading and speed changes

Cooperative decision framework

Simplified ship dynamics

Zhao et al27.

PPO

AIS data features, traffic density

Continuous action space

Real-world data validation

High computational requirements

Our approach

Enhanced DQN

Comprehensive state representation with relative metrics

Discretized speed and heading changes

Balanced reward function, improved performance in multi-ship scenarios

[To be discussed in limitations section]

Back to article page