RL Infinite Mario

For Introduction to Reinforcement Learning, we were given an option to work on the Reinforcement Learning 2009 Competition problems. One of the domain problems was "Infinite Mario", the goal is basically to get the Mario agent to learn how to play a "Super Mario Bros."-type game through experience. He gets rewards (e.g. for collecting coins or finishing level) or penalties (e.g. for getting killed by an enemy) by interacting with the environment.

Link to "Infinite Mario" domain description: http://2009.rl-competition.org/mario.php

Presentation and videos at Richard Gibson's (my partner for the project) site: http://sites.google.com/site/richardggibson/publications-and-presentations

Category:

Education

Tags:

• Infinite
• Mario
• Reinforcement
• Learning
• RL

[rcparts]

What is your observation abstraction? And what algorithm exactly? Tabular Q-Learning? Some function approximation? Do you have any document with details of your implementation?

I'm starting to work with RL but the results aren't that good until now!

Thanks in advance!

[nickabourisk]

Hi rcparts, I sent you a PM about a document with details.

[RyanfaeScotland]

@nickabourisk Hey there I know this is about a year old now but if you still have it and are still following this video can you forward those details on to me aswell.

I'm looking at something similar for my honours project, the original plan was to make an agent that can play mario using reinforcement learning until my professor burst my bubble by telling me it was already done (damn it) so now I'm looking at working on and improving past work in the field.

[nickabourisk]

@RyanfaeScotland no problem. Let me know how it turns out! I've edited the video description with a link to my project partner's website (it contains more details). Good luck!

[nickabourisk]

Thanks, not anywhere near as impressive as your's. We placed 2nd in the RL competition (my partner made it better by adding options and other things).

I believe that this required on the order of several hundred to a thousand iterations.

Good luck with your competition as well! Excited to see the results.

[nickabourisk]

The game and environment (stage/levels) are all set up already. We just get a set of observations about the world and get to choose actions for Mario (move right, run, and jump). Because the observations are so huge, we have to abstract it to something workable.

[robinba2342364]

This is a really nice effort, well done! How many iterationts were required before you arrived at that state?

Good luck with the competition!

[FusionNinjin]

constant jumping, slow forward moveing at any place even at the places where there are obvius ways to avoid enemys whiout even getting on same path as enemys, or skip them purely just by running, then falling above or even past enemys..

[billyjoeboucher]

Whoaaa thats awsum.

Did you make the stage layout, or are they made already, and you just put your Mario through it?