We wanted to understand how artificial intelligence operates, as well as observing its strengths and weaknesses. This could not be better exemplified by the algorithm we built which plays video games. We seek to show the way artificial intelligence learns from its mistakes in order to improve itself. We gave the name Artificial Idiot to the project, in part as a play on words over the acronym of artificial intelligence (AI), but more importantly to highlight the fact that in the present day artificial intelligence cannot operate on par with human interaction.
Disclaimer: We are absolutely and categorically against all insults directed at anyone, human or machine. Our use of Artificial Idiot is purely satirical and for educational purposes. No computers were offended in the making of this project.
What it does
The algorithm learns to play simple games. With raw images being the input, the AI would interact with the game in the same way a human would.
How we built it
Main techniques we used include Reinforcement Learning (RL) and Convolutional Neural Network (CNN). The code is written in Python and CNN is built by PyTorch.
Challenges we ran into
- Mastering the use of deep RL
- Defining the reward function of RL
- Tuning of CNN
- Learning to adapt the AI to different games
- Learning/relearning Python
- Technical problems with some of our aging laptops leading to numerous reboots and frequent visits to the coffee area to wind down
Accomplishments that we're proud of
- Programming went very smoothly
- Worked well together as a diverse, multi disciplinarian team
- We got to sit back and chill while watching our AI play Chrome's T-Rex Dino Game (aka the unable to connect game)
What we learned
We learnt a large array of different programming languages as well as familiarizing ourselves with creating websites thanks to the free domains kindly provided by get.tech and domain.com.
Best Domain Name Challenge
(Artificial Idiot(s), they're puns on AI!)
What's next for Artificial Idiot
We reaffirm our aspirations that artificial intelligence will improve its performance by learning from its mistakes. We would also push for compatibility with more complex video games. We hope that in the near future we can create a new algorithm which would show a significant improvement to the AI so poor T-Rex does not get so beat up!
Source Code: https://github.com/iamlxb3/RL_SIMPLE_GAME