Skip to content

Joint Graphs

Joint graphs are the shared representation that connects world modelling, opponent modelling, and shielding.

A joint graph represents public environment states and joint actions across all agents.

Main Artifacts

The active pipeline uses two central graph artifacts:

legal_joint_graph.pkl
env_transition_graph.pkl

legal_joint_graph.pkl contains exact legal topology only.

env_transition_graph.pkl contains the same topology plus learned environment transition probabilities and rewards.

Opponent probabilities are not saved as a separate base artifact. They are applied later when inducing the focal-agent MDP for a given opponent model.

Why Joint Actions

In multi-agent environments, the next state usually depends on all agents' actions.

A joint action records the combined action choice. The focal agent controls its own component, while the opponent model supplies probabilities for the other agents' components.

Focal MDP Induction

Shielding needs a single-agent view for the focal agent.

The graph induction step combines:

  • environment transition probabilities,
  • opponent action probabilities,
  • the focal action being evaluated.

The result is a focal MDP for a specific opponent-model level or mixture of levels. Sound value iteration runs on this induced structure.