www.braintank.ca

Category: Special Interest

Special interest projects of special importance.

  • Deep Q Learning and Self Play Beat Humans at Super Auto Pets

    Deep Q Learning and Self Play Beat Humans at Super Auto Pets

    Super Auto Pets is a popular online zero-sum extensive form game that emphasizes taking discrete actions to ensemble a cohesive team of pets to win.

    To better understand various principle related to Deep Q Learning, I designed and built a AI modet to complete agains humas in the game. My work:

    • Broke the challenge into sub-components to minimize cost of computation.
    • Reduced a large action space into one dramatically smaller, but maintaining all the actions available in the game.
    • Created a domain-specific search algorithm that increased performance from 3.6 to 5.5 points per game.
    • Oversaw optimization. Tested how hyperparameters of the model affected AI performance.

    I was successfully able to develop an AI that outperforms human players in diverse game scenarios.

    This video explains the project in detail.

  • Using Transformers to Generate Crossword Puzzles

    Using Transformers to Generate Crossword Puzzles

    For this project, I trained a set of transformers to generate crossword puzzles, akin to the New York Times crosswords.

    • Used Huggingface’s API to utilize transfer learning for the NL tasks.
    • Used beam search to create multiple board candidates that were evaluated at each step.
    • Puzzles are available to play at nickvinden.com/crossword.
    Play Now

    How to Play

    Select cells with your mouse and type them in with your keyboard. To get a new puzzle click “New Puzzle”. There are 3 different options for puzzle clue generation.

    1. GPT4: Best Model.
    2. GPT3.5: Good Model.
    3. T5: Miserable solving experience.

    To see if your answers are correct select “autocheck”.

    Give a go at cracking these completely AI generated crossword puzzles.

  • Comparing Instance Segmentation with Models Trained with and without GAN training samples

    Comparing Instance Segmentation with Models Trained with and without GAN training samples

    Note that this is a project that is currently under development.

    • Extension to paper Exploring and Classifying Beef Retail Cuts Using Transfer Learning (2022), by Abdallah Abuzaid et al.
    • The goal of this project is to reduce the cost of manual human image annotation.
    • Used a StyleGAN2 architecture to take a set of 10 beef cut image-annotation pairs, and increase the set of images from 10 to as many as needed.
    • I intend to measure under which conditions using a GAN to create more data is beneficial to instance segmentation.

  • Deep Counterfactual Regret Minimization in a Multi-AgentIncomplete Information Game of “Take 5”

    • Take-5 is a multi-agent repeating imperfect-information normal-form tabletop card game. Players aim to anticipate their opponents’ moves and strategize with the cards in their hand to optimize their outcomes and maximize utility.
    • Based on a paper “Deep Counterfactual Regret Minimization” by N. Brown et al. I adapted it to allow for the agent to accurately predict counterfactual regret vectors for this game.
    • With lack of better assessment, the agent was able to beat my group of friends who play the game very often.
  • Everybody has a Chess AI… Here’s Mine

    Everybody has a Chess AI… Here’s Mine

    Work was based on “mastering chess and shogi by self-play with a general reinforcement learning algorithm”.

    • Used a game matrix to map the state of the game, and a CNN to predict the utility of average moves.
    • To solve a sparse-reward problem, I started by optimizing the RL algorithm to take as many opponent pieces as possible, then to take as many pieces as possible without losing own pieces, and then finally optimized it to win the game.
    • Algorithm performed decently, but had clear strategic flaws. I was limited in compute for this project.