Explaining deep reinforcement learning decisions in complex multiagent settings: towards enabling automation in air...

Area of Science:

• Artificial Intelligence
• Operations Research
• Aerospace Engineering

Background:

• Complex, large-scale multi-agent systems present significant decision-making challenges.
• Enhancing human performance and engagement in task execution is crucial.
• Current Air Traffic Management (ATM) systems face demand-capacity imbalances requiring advanced solutions.

Purpose of the Study:

• To advance automation for decision-making in complex, large-scale multi-agent settings.
• To develop a deep multi-agent reinforcement learning method for resolving demand-capacity imbalances in Air Traffic Management (ATM).
• To provide high-quality solutions and high-fidelity explanations for automated decisions.

Main Methods:

• A deep multi-agent reinforcement learning approach was developed.
• The method enables agents to jointly decide on measures to resolve imbalances.
• Visual analytics tools were employed for rendering and exploring decision explanations.

Main Results:

• The proposed method addresses major challenges of scalability and complexity in ATM.
• Evaluation tests demonstrate the model's ability to provide high-quality solutions.
• The system generates high-fidelity explanations for the automated decisions.

Conclusions:

• The deep multi-agent reinforcement learning method effectively resolves demand-capacity imbalances in large-scale ATM settings.
• The approach enhances decision-making automation, improving human performance and engagement.
• The integration of explainability through visual analytics is a key contribution.