🚧 Xander: A D&D Combat Environment for Reinforcement Learning

Hello, this is a university project that attempts to make an RL environment for D&D 5E combat, where agents can play the role of monsters against players.

This is still very much in progress.

Features

Xander supports:

• Dice Notation:
  • XdY epxressions
  • Modifiers
  • Advantage/Disadvantage (2d20kl1, 2d20kh1)
• Stat Blocks:
  • Loading from JSON
  • Ability Scores, Modifiers, and Skills (with proficiencies)
  • HP, Temporary HP
  • Checks, Saving Throws
  • Death Saves (in progress)
• [ x] Visualization:
  • Combat arena map
  • _repr_html_ HTML display for Jupyter
• Reference Agents:
  • Q-Learning (in progress)

Building

Prerequisites

I still have to get this whole CI/CD thing figured out for Linux, but builds are available for MacOS and Windows.

To build from scratch, ensure you have:

• Rust toolchain: rustup, cargo with a nightly compiler (rustup install nightly)
• A new conda environment set up with Python 3.12 (conda create -n xander python=3.12)
• Maturin (cargo install maturin)
• Wasm-pack (cargo install wasm-pack)

Building the Project

• Run maturin build to build a Python .whl file.
• Run wasm-pack build -- --features web to build for wasm.

Implementation

• The core environment is written in Rust, and exposed to python via pyo3.
  • Python type bindings are available, but they do have to be manually updated (WIP).
• TODO: write more here.

Examples

from xander.engine import dice
from xander.engine.actors import Stats

dice.random_seed()

rat = State.from_json("tests/rat.json")
print(rat)

Explore Why this initial DQN didn't work.

• High variance in Q-values -> Dueling
• Spare Rewards -> Prioritized Replay Buffer.

Notes

Put the reward functions for your environments.

5 runs for the confidence interval.