Multi-agent reinforcement learning

Reward Shaping for Cooperative Multi-Agent Locomotion

A study of how reward shaping affects convergence and coordination in cooperative multi-agent locomotion.

Year
2024
Status
Project
Topics
Multi-agent RL · Reward design · Coordination
Several agents coordinating to carry a shared object

A conceptual illustration of the mathematical idea, not a plot of measured experimental results.

01

The question

Which reward signals teach independent agents to coordinate without destabilizing learning?

02

Central insight

A team reward expresses the final goal but can provide weak credit assignment. Shaped local signals can help, yet may create strategies that optimize the proxy instead of the task.

03

Approach

  1. 01

    Design reward formulations for cooperative locomotion.

  2. 02

    Train multi-agent policies using the proposed rewards.

  3. 03

    Benchmark the effect on convergence and coordination.

04

My contribution

  • Designed reward shaping for cooperative multi-agent reinforcement learning.
  • Benchmarked its effect on convergence and coordination.
05

Result

The evaluation examines how changing the reward affects learning behavior and coordination between agents.

06

What remains

A possible next investigation is how consistently the coordination behavior appears across independent training runs.