AIRIS is a General Machine Intelligence (GMI) project that combines aspects of Reinforcement Learning (RL) with more traditional symbolic techniques (GOFAI).
See http://airis-ai.com for videos and additional information.
Requires Python 3.x
See "requirements.txt" for additional required packages
Regardless of the test environment, AIRIS's learned knowledge is exported as one large python dictionary called "Knowledge.npy" whenever the save_knowledge() function is called. If this file exists, it is loaded on the startup of any environment. If this file is deleted or moved from the working directory, AIRIS will "start from scratch" and generate a new one.
puzzle_game_driver.py is a grid world puzzle game. Run it to watch AIRIS play the game.
Gif of AIRIS Completing the puzzle
puzzle_game_driver_custom.py This is the grid-world puzzle game, but loads the custom levels stored in \custom_levels instead of the original 13.
airis_stable.py is the AI code itself.
Knowledge.npy is the file the AI generates during runtime that has the knowledge it creates.
Knowledge_View.py Generates a text file of all the data contained in Knowledge.npy as knowledge_view.txt
rule_viewer.py outputs the knowledge into a more human readable format as rules_view.txt
minds_eye.py is the "Minds Eye" tool I used in the 2023 demo video. Good for seeing what the AI is planning to do and watching how it learns. It's a hacky tool put together by modifying the puzzle game code. It will only work with the puzzle game. Need to make one that works with any environment...
Knowledge - Trained Copy.npy is a copy of its knowledge after running through all levels 3 times. Rename it to Knowledge.npy and it will load it instead learning from scratch.
images folder stores the game graphics
logs folder stores debug output logs when self.debug (line 36 of airis_stable.py) is set to True. These can get quite large!
plan_log and predict_log folders are the where the AI outputs its plans and states that are read by minds_eye.py
screens folder is screenshots of each timestep.
constants.py settings used by puzzle_game_driver_universal.py
game_objects.py object data used by puzzle_game_driver_universal.py
state.py Used by the AI to store state information
minds_eye_constants.py settings used by minds_eye.py
minds_eye_objects.py object data used by minds_eye.py
AI Controlled - change line 1245 to true or false to toggle whether the AI is controlling the game or if it is observing a human player.
Plan Review Mode - change line 1246 to true or false to toggle whether the AI waits between plans for the user to allow it to continue. Can also be toggled during runtime.
Arrow Keys (Not AI Controlled) - Move the character
Space Bar (Plan Review Mode) - Allow the AI to continue
Space Bar (Not Plan Review Mode) - Pause the game
T - Toggle game speed
X - Safely exit the game. This ensures that the game doesn't close in the middle of the AI writing to Knowledge.npy which would corrupt it.
A - Toggle "Plan Review Mode"
Before using, clear the plan_log and predict_log folders. If you don't clear out the folders then minds_eye will be showing what a previous AI was thinking.
Run puzzle_game_driver_universal.py, then run minds_eye.py
M - Switch views between "Current Plan" and "Predicted States".
"Current Plan" shows the AI's most recent plan. "Predicted States" shows each state in the action/state graph used to create the current plan.
This is the level editor for the puzzle game environment. You can use this editor to create your own custom levels for AIRIS to solve. Run puzzle_game_driver_custom.py to use the custom levels.
Left Mouse Button selects an object from the object menu at the top and places a selected object in the level.
Right Mouse Button deletes objects in the level.
The "Clear" button will erase all objects in the level.
A level must have at least 1 robot and 1 battery to unlock the "Play" button. During play mode, arrow keys move and you can either click on the red button at the top or press Spacebar to return to the editor. You will automatically return to the editor when you solve the level.
To unlock "Save" you must play and solve your level.
Levels are saved in \custom_levels in ".csv" format. The name of the level is the level number and is automatically assigned upon opening the editor and is shown in the window title. For example, if there is already "1.csv" and "2.csv" then your new level will be "3.csv".
You can load a level for editing by adding the level number as a command line argument. For example:
$ python puzzle_game_editor.py 2 will load level 2 for editing.
You can only save whatever level was created or loaded on startup. To create or load another level, you must exit and restart the editor.
This is the grid-world puzzle game with the original 13 levels, but with all AIRIS dependencies stripped out. This environment is for use with any other AI algorithm.
It is designed to send 2 environmental arrays to an AI: A 20 x 15 grid of the game world filled with the integer IDs that represent the various game objects, and a 1 x 2 grid of integers that represent the number of Keys collected and Fire Extinguishers collected respectively.
--------------------------------------------
id Object
--------------------------------------------
0 floor
1 character
2 wall
3 battery
4 door
5 key
6 fire extinguisher
7 fire
8 1 way arrow right
9 1 way arrow left
10 1 way arrow down
11 1 way arrow up
12 open door
13 player character on top of right arrow
14 player character on top of left arrow
15 player character on top of down arrow
16 player character on top of up arrow
17 player character in open door
--------------------------------------------
(source: game_objects.py)
It expects one of 5 actions to be returned: 'up', 'down', 'left', 'right', or 'nothing'. It then updates the game environnent accordingly, and sends the updated game environment arrays to the AI so that the AI can observe the changes.
By default, it is set to be human controlled with the arrow keys. This can be changed by setting ai_controlled to TRUE on line 1136. Just make sure your AI can get, handle, and return the necessary values (See line 129 and line 157).
Number recognition using the MNIST hand-written character dataset.
This is intended to be a simple usage example. It will ask for a word, then ask what the label of that word is. It will try to predict what the label is for new words. Both the word(s) and the label can be anything. In the example below, I used "Name" and "Location" for labels, and some random names and locations I came up with.
Enter a word:
Robert
The AI doesn't yet have knowledge
What is the label for this word?
Name
This is not the outcome I predicted
Saving new knowledge ...
Enter a word:
House
The AI thinks the label for this word is: Name
What is the label for this word?
Location
This is not the outcome I predicted
Saving new knowledge ...
Enter a word:
Bill
The AI thinks the label for this word is: Location
What is the label for this word?
Name
This is not the outcome I predicted
Saving new knowledge ...
Enter a word:
Bob
The AI thinks the label for this word is: Name
What is the label for this word?
Name
Enter a word:
Building
The AI thinks the label for this word is: Name
What is the label for this word?
Location
This is not the outcome I predicted
Saving new knowledge ...
Enter a word:
Home
The AI thinks the label for this word is: Name
What is the label for this word?
Location
This is not the outcome I predicted
Saving new knowledge ...
Enter a word:
Housing
The AI thinks the label for this word is: Location
What is the label for this word?
Location
Enter a word:
Apartment
The AI thinks the label for this word is: Location
What is the label for this word?
Location
Enter a word:
John
The AI thinks the label for this word is: Location
What is the label for this word?
Name
This is not the outcome I predicted
Saving new knowledge ...
Enter a word:
Jim
The AI thinks the label for this word is: Location
What is the label for this word?
Name
This is not the outcome I predicted
Saving new knowledge ...
Enter a word:
James
The AI thinks the label for this word is: Location
What is the label for this word?
Name
This is not the outcome I predicted
Saving new knowledge ...
Enter a word:
Joe
The AI thinks the label for this word is: Name
What is the label for this word?
Name
Enter a word:
Studio
The AI thinks the label for this word is: Name
What is the label for this word?
Location
This is not the outcome I predicted
Saving new knowledge ...
Enter a word:
Studio Apartment
The AI thinks the label for this word is: Location
What is the label for this word?
Location
Enter a word:
Office
The AI thinks the label for this word is: Name
What is the label for this word?
Location
This is not the outcome I predicted
Saving new knowledge ...
Enter a word:
Office Building
The AI thinks the label for this word is: Location
What is the label for this word?
Location
Enter a word: