RaidEnv

MMORPG Simulator is an open-source and computationally accessible research platform that simulates a boss-fighting scenario of a MMORPG game. We aim to support a broad base of multi-agent and procedural content generation (PCG) research that would be impractical or impossible to conduct in other environments.

Getting Started

Environment Setup

1. Install the Unity Editor with version 2020.3.25f1
2. Install the Unity ml-agents from this repository

pip install --target $PATH_TO_INSTALL$ mlagents=0.28   

Usage

Workspace: /Project/Assets/ML-Agents/Examples/MMORPG

Game Description

The genre of this game is MMORPG (Massively Multiplayer Online Role-Playing Game) and implements a boss fighting scenario.
A dragon (boss) attacks multiple players from a distance to kill all, and players with various classes also use their skills to slay dragons.
There are several player classes: (add class description with their skills)

This is the key for playing in Unity play mode.

Key	Description
A, D	Turn Player 1
W, S	Move Player 1
1	Use Skill 1
2	Use Skill 2
K	Kill Player 1 (Set health to zero)
R	Reset Episode

Training Playtesting Agents

To train multi-agents, in addition to downloading and opening this environment, you will need to install the ML-Agents Python package. Follow the getting started guide for more information on how to use the ML-Agents trainers.

mlagents-learn /app/MA/{RL_HEURISTIC/RL_GENERALIZATION}/configs/{CONFIG_NAME}.yaml --env=/app/envs/MMORPG_Linux/MMORPG_Training.x86_64 --no-graphics --force --run-id=mmorpg1

You will need to use the MMORPG_training.unity scenes for training.
A configuration YAML (Experiment/MA/RL_GENERALIZATION/configs or Experiment/MA/RL_HEURISTIC/configs) for ML-Agents is provided. Depending on your computer's capabilities, you can uncomment and increase the number of environments (num_envs).

Reward Shaping

To produce the results in the paper, we used the default environment as it is in this repo. However, we also provide environment parameters to adjust reward functions and control the environment from the trainer. In multi-agent docs, you can find more details about state, action, and reward variables.

Parameters	Description
Shooting Reward	(default = +0.01) A reward given to an agent at every step for a shooting it use
Hit Reward	(default = +0.05) A reward given to an agent when it hits a dragon with their shooting skills
Win Group Reward	(default = +1.0) A reward given to the team if they kill a dragon

Training PCG Agent

You can generate the contents in the game by adjusting the value of the interface of each generator prefab or by substituting the source value in the OverallGenerator.

Customizable Features

Parameter Category	Parameter Type	Parameter Name	Data Type	Range	Description
Character Statistics	Status	Health Point	integer	[0, 1000]	Maximum number that agent can be dealing with the damage
		Manna Point	integer	[0, 100]	The resource that used when agent active a skill
		Spell Power	integer	[0, 100]	Reference number used when calculating the skill damage
		Movement Speed	float	[1, 2]	Move speed of the agent
	Attack	Power	integer	[0, 100]	Reference number used when calculating the melee attack damage
		Range	integer	[1, 10]	Maximum distance of the melee attack can be reached
		Speed	float	[1, 2]	Number of melee attacks can be activated per second
	Defensive	Armor	integer	[0, 100]	Ratio that agent dealt reduced damage
		Evasion	integer	[0, 100]	Probability that agent can dodge melee attack or skill
		Parry	float	[0, 100]	Probability that agent can block melee attack or skill
	Primary	Strength	integer	[1, 100]	Measure of physical power and proficiency
		Agility	integer	[1, 100]	Measure of swiftness
		Intelligence	integer	[1, 100]	Measure of mental power and magical prowess
	Secondary	Critical	integer	[0, 100]	Chance that agent ability can be more effective than normal
		Haste	integer	[0, 100]	Determine the agent's action swiftness
		Versatility	integer	[0, 100]	Increase the agent's attack and defensive abilities effect
		Mastery	integer	[0, 100]	Enhance some abilities effect to emphasize character class
Skill	Information	Name	string	-	Name given to each skill
		Trigger Type	enum	[0, 1]	Determine the skill is active skill or passive skill
		Magic School	enum	[0, 9]	Determine the visual effects of the skill (e.g. fire, frost)
		Hit Type	enum	[0, 1]	Determine if the skill is a melee attack or skill
		Target Type	enum	[0, 2]	Type of target which the skill can be used on
		Projectile Speed	float	[0, 50]	Speed of the projectile that reaches the target
		Affect on Ally	bool	True/False	Determine if an agent can activate a skill on an ally
		Affect on Enemy	bool	True/False	Determine if an agent can activate a skill on an enemy
	Condition	Cool Time	float	[0, 60]	Amount of time required to activate the skill again
		Cast Time	float	[0, 2]	Amount of time required to execute the skill after activation
		Cost	float	[0, 100]	Amount of Mana required to activate the skill
		Range	float	[1, 20]	Maximum distance of the skill can be reached
		Charge	integer	[1, 3]	Number of skill activations regardless of the Cool time
		Cast on moving	bool	True/False	Determine if an agent can activate a skill during movement
		Cast on Casting	bool	True/False	Determine if an agent can activate a skill during casting
		Cast on Channeling	bool	True/False	Determine if an agent can activate a skill during executing
	Coefficient	Value	float	[0, 2]	Scaling factor that determines the damage of the skill