tetris-ai

A bot that plays tetris using deep reinforcement learning.

Demo

First 10000 points, after some training.

How does it work

Reinforcement Learning

At first, the agent will play random moves, saving the states and the given reward in a limited queue (replay memory). At the end of each episode (game), the agent will train itself (using a neural network) with a random sample of the replay memory. As more and more games are played, the agent becomes smarter, achieving higher and higher scores.

Since in reinforcement learning once an agent discovers a good 'path' it will stick with it, it was also considered an exploration variable (that decreases over time), so that the agent picks sometimes a random action instead of the one it considers the best. This way, it can discover new 'paths' to achieve higher scores.

Training

The training is based on the Q Learning algorithm. Instead of using just the current state and reward obtained to train the network, it is used Q Learning (that considers the transition from the current state to the future one) to find out what is the best possible score of all the given states considering the future rewards, i.e., the algorithm is not greedy. This allows for the agent to take some moves that might not give an immediate reward, so it can get a bigger one later on (e.g. waiting to clear multiple lines instead of a single one).

The neural network will be updated with the given data (considering a play with reward reward that moves from state to next_state, the latter having an expected value of Q_next_state, found using the prediction from the neural network):

if not terminal state (last round): Q_state = reward + discount × Q_next_state else: Q_state = reward

Best Action

Most of the deep Q Learning strategies used output a vector of values for a certain state. Each position of the vector maps to some action (ex: left, right, ...), and the position with the higher value is selected.

However, the strategy implemented was slightly different. For some round of Tetris, the states for all the possible moves will be collected. Each state will be inserted in the neural network, to predict the score obtained. The action whose state outputs the biggest value will be played.

Game State

It was considered several attributes to train the network. Since there were many, after several tests, a conclusion was reached that only the first four present were necessary to train:

• Number of lines cleared
• Number of holes
• Bumpiness (sum of the difference between heights of adjacent pairs of columns)
• Total Height
• Max height
• Min height
• Max bumpiness
• Next piece
• Current piece

Game Score

Each block placed yields 1 point. When clearing lines, the given score is number_lines_cleared^2 × board_width. Losing a game subtracts 1 point.

Implementation

All the code was implemented using Python. For the neural network, it was used the framework Keras with Tensorflow as backend.

Internal Structure

The agent is formed by a deep neural network, with variable number of layers, neurons per layer, activation functions, loss function, optimizer, etc. By default, it was chosen a neural network with 2 hidden layers (32 neurons each); the activations ReLu for the inner layers and the Linear for the last one; Mean Squared Error as the loss function; Adam as the optimizer; Epsilon (exploration) starting at 1 and ending at 0, when the number of episodes reaches 75%; Discount at 0.95 (significance given to the future rewards, instead of the immediate ones).

Training

For the training, the replay queue had size 20000, with a random sample of 512 selected for training each episode, using 1 epoch.

Requirements

• Tensorflow/Jax/PyTorch
• Tensorboard
• Keras
• Opencv-python
• Numpy
• Pillow
• Tqdm

The original tests were evaluated using Keras with Tensorflow as the backend. However, using Jax as the backend appears to result faster training/predict operations.

Run

• Train/execute:

# hyper parameters can be changed in the run.py script
python3 run.py

• View logs with tensorboard:

tensorboard --logdir ./logs

Results

For 2000 episodes, with epsilon ending at 1500, the agent kept going for too long around episode 1460, so it had to be terminated. Here is a chart with the maximum score every 50 episodes, until episode 1450:

Note: Decreasing the epsilon_end_episode could make the agent achieve better results in a smaller number of episodes.

Useful Links

Deep Q Learning

• PythonProgramming - https://pythonprogramming.net/q-learning-reinforcement-learning-python-tutorial/
• Keon - https://keon.io/deep-q-learning/
• Towards Data Science - https://towardsdatascience.com/self-learning-ai-agents-part-ii-deep-q-learning-b5ac60c3f47

Tetris

• Code My Road - https://codemyroad.wordpress.com/2013/04/14/tetris-ai-the-near-perfect-player/ (uses evolutionary strategies)