Home

Welcome to SusNet!

This wiki compiles a growing list of essential resources for the SusNet project, aimed at training agents for optimal roles in both Imposter and Crew member scenarios within a zero-sum game environment. The project starts with a fully observable world and plans to incrementally introduce more complexity by decreasing observability.

Defining Multi-Agent Environments:

To support the creation of environments that accommodate multiple agents, these resources are recommended:

• MAgent2: Provides efficient implementations of grid world environments for pixel-sized agents. It helps understand the influence of environments on agent interactions in multi-agent setups. Specifically, the grid_world.py script is notable for its approach to managing events, agents, and the environment, offering insights valuable for multi-agent problem-solving.

Exploring Control Methods:

Here are some resources for exploring Control Methods in a Adversarial Multi-Agent Dynamic.

• Smoothing Policy Iteration for Zero-sum Markov Games: This paper introduces a strategy for policy iteration in zero-sum Markov games, moving beyond the constraints of traditional min-max algorithms. It stands out for two main reasons:

  1. It provides insights into multi-agent environment dynamics and decision-making processes in zero-sum games, closely aligning with our project's scenarios.
  2. It details a theoretically solid algorithm for policy iteration that reliably converges, covering both tabular and functional approximation implementations.

• Multi-Agent Learning: Rollout and Policy Iteration General Policy Iteration application for Multi-Agent problems. Big Limitation: setup requires that problem is fully observable and order of agent operations is fixed and sequential. Uses: One-step Lookahead Rollout (need to look into what this means exactly). However, this is a DP method.