This repository is my playground for reinforcement learning.

Project 1.0: Movie Recommender with (DDQN and GNN)

Descriptions

• All files and training notebooks are in the Movie_Recommender folder
• Data Set: MovieLens 100K
• The Baseline model: Double DQN with STOCHASTIC PRIORITIZED REPLAY
  • Selected Features: movie title, movie date, genre, movie rating avg, user age, user average rating (for all and each genre), user occupation
  • Output: Predicted rating of the movie.
  • Metrics: MSE Loss.
  • DQN architecture:
    • [FC1] -> relu -> [FC2] -> relu -> [FC3] -> log_softmax -> output
  • Game Setup:
    • State = Concatnated embedding of selected features
    • Action space = Discrete(5). Actions 0-4 are correspondent to rating 1-5
    • Rewards:
      • D = abs(rating - 1 - action)
      • D = 0 -> +3
      • D = 1 -> +1
      • D = 2 -> -1
      • D >= 3 -> -3

Results

-Baseline Model: RMSE = 1.18 after 10 epochs of training.

Project 1.1: Sequential Recommendation based on RL (TB finished)

Project 2: Q-learning for Hangman.

• All files and training notebooks are in the hangman folder.
• This repository provided gym environments and agents for both table-based and Deep NN Q-learning.
• State:
  • observation 0: the state of current word 0-25 <=> a-z, and 26 = "_".
  • observation 1: guessed letters. Binary(26). 1: guessed 0: not guessed
  • observation 2: attempts left - 0 to 6
• Action: choose between a-z, encoded as 0-25
• Reward Rule:
  • correct guess +3
  • wrong guess -1
  • win game +10
  • lose game -10
• Result after one epoch on table-based Q-learning: 61%