The research group Adversarial Resilience Learning (ARL) researches agents based on deep reinforcement learning to resiliently operate critical infrastructures. The ARL group's approach includes learning from expert knowledge via offline learning/imitation learning, as well as an approach towards explainability (XRL). 

The ARL group usually verifies its approaches by simulating power grids. However, the modern DRL community readily accepts games as universal benchmarks, as they are easy to understand for a wide audience, yet exhibit challenging properties for DRL algorithms. However, variants for autocurricula, especially those where agents do not simply take opposide roles with conceptionally similar sensor/actuator sets, are still missing.

Goal of this thesis is to create an environment for autocurriculum learning. The environment should be in the spirit of the game "Plague, Inc.," in that it simulates the outbreak of a Zombie Virus and the world's reaction to it. In this environment, two types of agents take part (hence, the autocurriculum setting):

1. The “attacker,” i.e., an agent that creates and mutates a zombie virus,
2. the “world,” represented by an agent that can take the usual actions of shutting down airports/trade routes, closing boarders, researching remedies, etc. to stop the spread.

The thesis can be either a bachlor or master thesis. If pursued as bachelors thesis, the following objectives should be met:

• Creation of an environment with rudimentary, yet correct, simulation mechanics of a world, spreading of a virus, mutation possibilities of that virus, etc.
• Interfaces for the two types of agents
• Visualization of agent performance, zombie virus state, and world state; desireably in Grafana

If prusued as master thesis, additionally, the following objectives should be considered:

• Training of DRL agents
• Visualization of XRL features, based on the ARL architecture

You should have solid knowledge in Python application development. Having prior knowledge in Grafana or any other web-based visualization tool is a plus. Knowledge of Deep Reinforcement Learning algorithms is beneficial, but not required.

The thesis can be written in English or German.