Self-Adaptive Swarm Systems (SASS) -- Abstract

Multi-agent systems (MAS) could play a pivotal role in realizing future intelligent workspaces, especially in building so-called artificial social systems, such as self-driving cars and multi-robot systems (MRS). For example, MAS/MRS cooperates to increase mission performance in many applications, including exploration, surveillance, defense, humanitarian, and emergency missions like urban search and rescue (USAR). In such missions, complex environments such as hazardous, dynamic changing, and adversarial surroundings create a significant challenge to the agents in realizing their full potential. Therefore, this thesis addresses some pressing gaps in the literature in realizing an adaptive MAS by proposing a principled MAS cooperation framework, termed the Self-Adaptive Swarm System (SASS), which bridges communication, planning, decision-making and learning in the distributed MAS.

ProQuest Version Dissertation: Self-Adaptive Swarm Systems (SASS)

Full Version Download: Self-Adaptive Swarm Systems (SASS)

Cite the Dissertation:

@phdthesis{yang2022self,
  title={Self-Adaptive Swarm System},
  author={Yang, Qin},
  year={2022},
  school={University of Georgia}
}

Contributions

The core scientific contributions of this thesis are as follows:

• 1. We define a novel human-inspired Agent (robot) Needs Hierarchy model to consider an agent’s motivation and requirements based on the current status and assigned tasks;
• 2. We present a priority-based distributed Negotiation-Agreement Mechanism for realizing multi-agent tasks assignment problems, effectively avoiding plan conflicts – Here, we decompose the tasks into Atomic Operations and achieve MAS cooperation through a series of simple sub-tasks;

*Note: Check the Link for more details.

• 3. We introduce a new needs-based agent trust and cooperation mechanism – Relative Needs Entropy (RNE) – to create needs-driven relationships among multiple agents in challenging environments;

The simulation of two heterogeneous robot teams cooperative achieving tasks in USAR with Unity:

*Note: Check the Link1 and Link2 for further reading.

• 4. We build a new hierarchical utility network – Game-theoretic Utility Tree (GUT) – to realize game-theoretic solutions for the cooperating MAS in the presence of adversarial agents;

The simulation of explorers against adversaries with GUT achieving a task in Explore Domain:

The Explore Domain in Robotarium: Greedy Approach vs Random Selection vs GUT

*Note: Check the Link for more details.

The Pursuit Domain in Robotarium: Constant Bearing (CB) vs Pure Pursuit (PP) vs GUT

*Note: Check the Link for more details.

• 5. We propose a novel Bayesian Strategy Networks (BSN) applied to deep reinforcement learning by decomposing tasks into multiple sub-level actions and obtaining the optimal agent policies in unknown and challenging environments.

Demonstration in MuJoCo with OpenAI Gym: Hopper-v2, Walker2d-v2, Humanoid-v2

*Note: Check the Link for more details.