maplegym adds MapleStory to OpenAI Gym so you can teach programs/neural networks to play for you.

pip install gym
git clone https://github.com/brhs/maplegym

Move the maplegym folder into your project directory or $PYTHONPATH so you can import it. Install maplegym's modified MapleStory client.

First, read the Gym documentation.

Running import maplegym registers maplegym as an environment with Gym. Create an environment using gym.make('MapleGym-v0').

To slice gameplay into timesteps, we define a "step" as a 0.2s pause after each action. Change this by setting maplegym.conf.timestep.

Here are the standard Gym variables:

• obs (600×800×3 uint8 array) is the game screen's RGB pixel data.
• action (uint8) is a number in range(20); see this.
• reward (uint32) is how much your cumulative EXP plus money has increased in the last timestep.
• done (bool) is True if you've reached level 70.
• info (dict) usually contains diagnostic info but maplegym doesn't use it.

maplegym uses an old version of the game (v.55) from about 2008. The maplegym Env automatically starts a (very) small local MapleStory server and runs the game client.

• You're a Bandit.
• The game takes place on a small set of maps; when you reach the right level, you're automatically teleported to the next map.
• Items, AP, and SP are disabled. Your stats and skills are raised automatically, you automatically get new eqiups, and when you pick up items you're immediately compensated for the item's shop price.
• There's no HP or MP. When you're damaged or use a skill, you lose mesos equal to the amount of HP or MP lost. (Your mesos can't go under 0.)
• The game UI is completely removed.
• All keys and mouse clicks are disabled except Control, Z, Alt, and the arrow keys, which correspond to Attack, Loot, Jump, and movement. Buffs are autocast, and "attack" is automatically mapped to your latest feasible attacking skill (regular attack, Double Stab, or Savage Blow). [1]
• The EXP curve is rewritten to be much more lenient, since the goal is not an environment that takes forever to beat.

This reduces the game to a bitmap state, a small set of abstract actions, and a unified reward structure.

Many reinforcement learning algorithms start with a seed of past experiences so the agent doesn't spend ages stumbling around like a human toddler. The maplegym.recorder.record function lets you play the game yourself and record your gameplay:

import maplegym

for obs1, action, reward, obs2 in maplegym.recorder.record():
    # record the experience here
    pass

maplegym.recorder.record opens a MapleStory window and yields an experience 4-tuple after every timestep. These mean the same thing as described described above in Usage. You can save them to a file or whatever.

The next milestone is to make maplegym extensible and implement everything in code instead of mysteriously binary patches that no one understands. There's a lot more documentation to be written.

Progress largely depends how much attention this gets (I'm not hopeful), how motivated I feel, and whether the capitalist grim reaper comes calling again soon. Pull requests and issues are warmly welcomed.

[1] Actually, buffs are not currently autocast. There's about 20 lines of code left to get it working. Someone open an issue to remind me.