snake ai

running files

just cd to the directory you want. and run python ./main.py(uncomment or comment the train and test function)

info

the prefixes growing and no_growing at each of each directory is just meant to reduce the complexitiy of task. (if no_growing then when snake eates food it won't increase it's length)

in each directory:

• agent.py: well this is the agent.(a that player learns to play game). and the type of it is mentioned on the suffixes (e.g. q_learning)

• helper.py: the helper functions(e.g. for plotting the results)

• main.py: has 2 functions train and test

• snake.py: the game(enviroment). i've implemented it, and i tried to make it like gym enviroments

• state_setting.py: the function used to implement the inputs that the agent will receive from enviroment (that makes the game more flexible)

directories

no_growing_snake_q_learning

after about less than 20_000 training steps the score passes from 1000 (Note the q values has been replaced so it might now get stuck in some places)

growing_snake_q_learning

well as the chart suggests, the agent is unable to go further than 50-60 because of the lack of information. (or maybe tweaking the hyper parameters might help)

growing_snake_deep_q_learning

well, the algorithm works fine if i've trained it more, (+ wider VISION_SIZE) probably it could master the game.

there is agent_cpu.py and agent_gpu.py. if you have a gpu access for tensorflow you probably wana use agent_gpu. if not you should use agent_cpu (the only thing that makes computations very expensive in agent_gpu is tf.queue.RandomShuffleQueue. for some reasons tf.queue.RandomShuffleQueue uses a lot resources in cpu).

and note that while useing snake.py the width and height should be equal. otherwise the error will raise from state_setting.py

the file state_setting.py know returns a window of the snake head(25)* + direction of food(4) + direction of snake(4).

where:

• 0: is empty space
• -1: is the food
• 1: is snake head
• 2: is snake body
• 3: is walls

evaluate_agent.py this is used for testing the agent with several hyperparameters in cartpole enviroment. the AdamW and xaiver init method worked well(learned faster)