Algorithms and Baselines¶

This page names the main algorithms and baselines that appear in notebooks and outputs.

IPPO¶

IPPO is independent proximal policy optimization. Each agent is trained with a decentralized policy update.

In this repo, IPPO is the common baseline learner and the base policy used for shielded variants.

IPPO Lagrangian¶

IPPO Lagrangian adds a learned penalty for safety violations or constraint costs.

It is a useful comparison for shielded control because it encourages safer behavior through training loss rather than replacing actions at runtime.

ICPO¶

ICPO is the repo-native constrained policy optimization baseline.

It provides another comparison point for reward-constrained and safety-aware learning.

Learned Shielded IPPO¶

Learned shielded IPPO trains with TransitionShield around the environment.

The learner still proposes actions from the base action space. The shield may replace unsafe proposed actions based on learned world-model and opponent-model artifacts.

True Shielded IPPO¶

True shielded IPPO is an exact-transition baseline where available.

It is useful for understanding how much of the learned shield's behavior comes from modelling error versus the shielding method itself.