Mario AI - a comparison

This is a Mario AI that I and a friend wrote for an AI course we took in Fall 2009 at UNC-Charlotte. In this video I explain about a few reasons why our AI is slightly different and arguably more legitimate than Robin Baumgarten's implementation. We did not submit this to the Mario AI competition but instead took that competition's rules and used them as the outline for a final AI project in our class.

Robin's implementation is great to watch but as you will see in the video we feel his method is a little sketchy, though he did win the competition. Congratulations to him and we look forward to next year's competition, or similar events.

Lastly, I apologize for the video quality of the Mario game, the Infinite Mario output screen is rather small and of low visual quality, unfortunately this is the best I can get the quality to be.

EDIT: A few errors in this video have been brought to my attention and so I am currently reviewing them and will edit as necessary. Thank you to those in the Mario AI community for your help and I appreciate your contributions, I want to provide as accurate information as possible.

Category:

Education

Tags:

• Mario
• AI
• Artificial
• Intelligence

[nemetroid]

I believe you explain RobinB's implementation in the wrong way. As I understand it, he is taking the grid data and feeding it into a procedure for creating a level environment, which he is then feeding into his copy of the engine.

You make it come across as if he is reading the game state directly. This is not the case, take a look at his LevelScene.java and you'll see that he is parsing the data from the environment interface.

[melink14]

You realize the early in the competition they released float values for all enemies as part of the observation?

[QuadmanSwe]

This was awesome!

[Drecon84]

Time for random levels?

[Gargomon251]

why is that fraps video always impossible to see?

[TheSpiderZen]

why not make more squares than 22? should fix the problem with "is it in this square or that one?" unless that is the challenge of the whole contest and i lost that detail in the nerd speek?

[sabriath]

I was under the impression that he is using pathfinding to the END OF THE SCREEN....the only information he uses from game mechanics is knowledge of the speed of enemies. That means as soon as the enemy appears, the pathfinding algorithm knows the location the enemy will be from that point on. When the enemy is gone, the algorithm no longer "sees" it. How is that cheating? Mario "sees" what we "see", we as humans can see speed of enemies. Whatever though.

[Phonolith303]

Yes, Mario isn't cheating.

Some (real pro) players sometimes play like that.

[jagwio]

@nemetroid good to know

[desposyy]

this is probable the single geekiest and most interesting thing i have ever seen. thankyou