Projects

Rhymebot3000 is a TTS ML model capable of rapping on a given bpm.  It is based on SotA transformer architecture with over 34h of rap data for rhythm prediction.  Check out the website for more information.

Disclaimer: The generated videos from Rhymebot3000 are intended for research purposes only, is based on algorithms, and does not represent the opinions, thoughts, or intentions of any real person.

Kuhn Poker is a simplified poker game with two players and three cards, for example J, Q, K represented as 1, 2, 3 in the program and only one betting round.  Through self-play, the optimal playing strategy is realized in the limit.  The optimal strategy is determined through counterfactual regret minimization algorithm that approximates the Nash equilibrium. 

Regret is a number that describes how much one regrets a particular decision. Some examples are: folding a good hand (wrong thing to do, high regret), folding a bad hand (right thing to do, low regret).  The strategy in a particular position is averaged over the number of iterations based on this regret and therefore gives the optimal strategy in any particular position.  This theory can be applied to more complex poker games like NLH, LH, etc.

Rules:

Basically limit hold'em (bet of 1) with one betting street and ante of 1.  E.g.:

• Both players ante 1 into pot and are each dealt one of the three cards

• Player 1 has option to check or bet 1

  • If Player 1 bet:

    • Player 2 can call or fold

  • If Player 1 check:

    • Player 2 can bet, or check

      • If Player 2 bet:

        • Player 1 can call or fold

Tree notation:

1 - J
2 - Q
3 - K

The first column of numbers is the percentage that should check or fold.

The second column is the percentage that should bet or call.

For example, 1cb, means that in the position that player 1 checked, and is facing player 2 bet, he should always (0.999999) check/fold (in this case fold is the only legal action) and never call.  Since J loses to Q or K, one should never go to showdown and lose more than necessary.

Note that Fig.1 & Fig.2 show slightly different strategies for certain states (i.e. 1, 2cb, 3), all for the first player, while all states associated with second player are fixed.  It has been demonstrated that there are infinite amount of equilibrium strategies for the first player making this a mixed-strategy Nash equilibrium. 

Code: Link

 Sample game

Note:

- Game begins as white only

- Grid is (0,0) at top left, (7,7) at bottom right

- No checking, pawn promotion, castling, en passant

- Game ends when king is taken 

- Minimax algorithm implemented with alpha-beta pruning, searches 5 moves deep 

- Search tree optimized by implementing MVVLVA (Most Valuable Victim, Least Valuable Attacker) i.e. sort moves to prune faster

Code: Link