Benjscho / mdptetris-experiments

Repository for dissertation agent experiments

MdpTetris experiments

This repository holds the agents and experiments to be run on gym-mdptetris.

Installation

This repository can be installed via GitHub:

pip install git+https://github.com/Benjscho/mdptetris-experiments

Or by cloning to the desired directory, cd-ing into the directory and running pip install -e ..

Run experiments

Below are commands to get started with running the agents provided. Further commands used can be found in ./mdptetris_experiments/experiments/experiment-commands.md.

Linear agent

The linear agent can be run with the default weights created by Pierre Dellacherie using python mdptetris_experiments/agents/linear_agent.py render. This agent is very effective at clearing lines, so this can take some time to run.

This agent is very successful:

To customize the weighting of the agent, a new instance can be created:

import numpy as np
from mdptetris_experiments.agents.linear_agent import LinearGame

agent = LinearGame(weights=np.array([-1, 1, -1, -1, -4, -1]))
cleared, duration = agent.play_game()
print(f"{cleared:,} rows cleared")

MBDQN

The model-based DQN agent utilises a similar approach to the linear agent, generating all of the possible subsequent states before using a neural network to evaluate the value of each board state. This agent performs at a poor level but demonstrates learning after 3000 epochs.

To train the agent with default settings simply run:

python mdptetris_experiments/agents/MBDQN/train.py

A trained agent can then be tested by providing it with a model load file:

python mdptetris_experiments/agents/MBDQN/train.py --test --render --load_file <LOAD_FILE_PATH>

The agent has a number of hyperparameters that can be varied via input from the command line. You can view these by running the program with the -h or --help parameter:

python mdptetris_experiments/agents/MBDQN/train.py -h

For example, to run the one piece per episode training experiments, where each episode only drops one of the 7 Tetris pieces, you can use the --one_piece option. To specify which GPU to use while training use the option --gpu=<GPU-ID-HERE>. To vary the board height, use --board_height=10.

PPO

The PPO-Clip agent is a model-free agent that learns through interaction with the environment. To train the agent with default settings run:

python mdptetris_experiments/agents/PPO/train.py

Similarly to the MBDQN agent, there are a number of hyperparameters that can be varied through command line options. You can see the available options by running with the -h parameter:

python mdptetris_experiments/agents/PPO/train.py -h