Quack-TD

• Usage
  • Evaluation methods
  • Examples
    • Train default model
    • Train model named quack
    • Train default model in sessions of 10,000 episodes
    • Train model quack and evaluate after each training session
    • Evaluate model named quack using default evaluation method (currently random)
    • Evaluate default model using evaluation methods random and pubeval
• Model storage format
  • Files
  • Log format
    • Evaluation log (eval.log)
    • Training log (train.log)

Quack-TD is a backgammon playing algorithm based upon neural networks trained through TD(λ)-learning. The algorithm is implemented using Python 3 and Tensorflow.

Usage

The main executable is main.py. Various command-line options and switches can be used to execute different stages and modify the behaviour of the program. All command-line options and switches are listed by running main.py with the argument --help. The three central switches are listed below:

• --train: Trains the neural network for a set amount of episodes (full games of backgammon) set by --episodes (defaults to 1,000). Summary results of the games played during the training session are written to models/$MODEL/logs/eval.log
• --eval: Evaluates the nerual network using the methods specified by --eval-methods for a the amount of episodes set by --episodes (defaults to 1,000). Results are written to models/$MODEL/logs/eval.log.
• --play: Allows the user to interactively play a game of backgammon against the algorithm.

Evaluation methods

Currently, the following evaluation methods are implemented:

• pubeval: Evaluates against the pubeval backgammon benchmark developed by Gerald Tesauro. The source code is included in the pubeval directory and needs to be compiled before use. The binary should be placed at pubeval/pubeval.
• random: Evaluates by playing against a player that makes random moves drawn from the set of legal moves. Should be used with high episode counts to lower variance. TODO: Doesn't even work currently

Examples

The following examples describe commmon operations.

Train default model

python3 --train

Train model named quack

python3 --train --model=quack

Train default model in sessions of 10,000 episodes

python3 --train --episodes=10000

Train model quack and evaluate after each training session

python3 --train --eval-after-train --model=quack

Evaluate model named quack using default evaluation method (currently random)

python3 --eval --model=quack

Evaluate default model using evaluation methods random and pubeval

python3 --eval --eval-methods random pubeval

Model storage format

Models are stored in the directory models. If no model is specfied with the --model option, the model is stored in the models/default directory. Otherwise, the model is stored in models/$MODEL.

Files

Along with the Tensorflow checkpoint files in the directory, the following files are stored:

• episodes_trained: The number of episodes of training performed with the model
• logs/eval.log: Log of all completed evaluations performed on the model. The format of this file is specified in /Pownie/backgammon/src/commit/554e587ffde3e60d72aa2dd64f8cb2cdccf9cbf8/Log%20format.
• logs/train.log: Log of all completed training sessions performed on the model. If a training session is aborted before the pre-specified episode target is reached, nothing will be written to this file, although episodes_trained will be updated every time the model is saved to disk. The format of this file is specified in /Pownie/backgammon/src/commit/554e587ffde3e60d72aa2dd64f8cb2cdccf9cbf8/Log%20format.

Log format

The evaluation and training log files (logs/eval.log and logs/train.log respectively) are CSV-foramtted files with structure as described below. Both files have semicolon-separated columns (;) and newline-separated rows (\n).

Evaluation log (eval.log)

Columns are written in the following order:

• time: Unix time (Epoch time) timestamp in local time (TODO: should be UTC instead?) describing when the evaluation was finished.
• method: Short string describing the method used for evaluation.
• trained_eps: Amount of episodes trained with the model before evaluation
• count: Amount of episodes used for evaluation
• sum: Sum of outcomes of the games played during evaluation. Outcomes are integers in the range of -2 to 2. A sum of 0 indicates that the evaluated algorithm scored neutrally. (TODO: Is this true?)
• mean: Mean of outcomes of the games played during evaluation. Outcomes are integers in the range of -2 to 2. A mean of 0 indicates that the evaluated algorithm scored neutrally. (TODO: Is this true?)

TODO: Add example of log row

Training log (train.log)

Columns are written in the following order:

• time: Unix time (Epoch time) timestamp in local time (TODO: should be UTC instead?) describing when the training session was finished.
• trained_eps: Amount of episodes trained with the model after the training session
• count: Amount of episodes used for training
• sum: Sum of outcomes of the games played during training. Outcomes are integers in the range of -2 to 2. A sum of 0 indicates that the evaluated algorithm scored neutrally. (TODO: Is this true?)
• mean: Mean of outcomes of the games played during training. Outcomes are integers in the range of -2 to 2. A mean of 0 indicates that the evaluated algorithm scored neutrally. (TODO: Is this true?)