MTSP with Reinforcement Learning Approaches

Environments

This environment built base on the gymnasium environment. Examples are shown on the environment creation documentation. The modified environment include:

• MTSPEnv(num_agents,num_tasks,map_boundary): Multiple Traveling salesman Problem
  • State: is a concatenated vector of agent position and a binary vector of remaining tasks
  • Action space: $a_{ij} = i*(\text{task number})+j$ means assign task $i$ for agent $j$
  • Transition probabilities: this is a deterministic environment, so $P({s}'|s,a) = {1, 0}$

Models

• Deep-Q Network The DQN code is implemented based on the examples that are shown on youtube video.

Results

The following plot visualizes the final solution obtained using

• DQN