BEP_How_To_Train_Your_Dragon

Welcome,

This code is part of the Research Project 21/22 Q4 at TU Delft of M.R. Tromp, Q.B. Hofstede (add your name here). The topics of the projects are:

Author	Project Title
M.R. Tromp	Alternatives to the components of a Genetic Algorithm
Q.B. Hofstede	How to train your dragon: on the application of the Metropolis-Hastings algorithm for Program Synthesis
Add your name	and your project here :)

The results and data processing methods from these papers can be found in the Published-Results-Q4-2022 folder.

This codebase is an extension of the codebase provided for this research, which is called BEP_project_synthesis.

BEP_project_synthesis

This code is part of the BEP projects of F. Azimzade, B. Jenneboer, N. Matulewicz, S. Rasing and V. van Wieringen.

Source of training/test data

Robot and Pixel test/training data was generated by A. Cropper and S. Dumančić for their paper "Learning large logic programs by going beyond entailment." arXiv preprint arXiv:2004.09855 (2020).

The String test/training data was received from S. Dumančić who took them from the paper by Lin, Dianhuan, et al. "Bias reformulation for one-shot function induction." (2014).

---

How to run:

Option 1: main.py
You can run the main.py file to run experiments. It contains comments with details on how to configure the experiment. e.g. which domain(s), which search algorithms etc.

For the Metropolis-Hastings algorithm you can also specify the domain as well as experiment parameter by calling main through the command line. For example:

python main.py "robot" "{'add_if_statement_random': 10,'add_loop_random': 10,'add_token_random': 10,'alpha': 1,'remove_token_random': 20,'start_over': 1,'type': 'metropolis_hastings'}"

You can of course store these parameters in a text file and generate it automatically with a script, for example through...

Option 2: DelftBlue
If you have access to the DHPC run DelftBlue super computer then you can create a series of SLURM jobs through the create_jobs.py file. It also contains details exactly on how to customize the various experiments it.

Results

The results will be output to the /results/ folder. This can be changed in the search/batch_run.py file's init_store_system method.

For example, this includes the commands to save the results to your scratch drive on DelftBlue (At the time of writing).

---

Genetic Algorithms

The code for this project is located in search/gen_prog/vanilla_GP_alternatives. The original code for VanillaGP is called vanilla_GP.py.

To run the alternatives code, make sure that main.py contains [VanillaGPReworked, "gpr"] in the list of search algorithms. If you want to change which alternatives are used, change _type to the wanted alternative. Any other settings were kept the same as in the original VanillaGP, but can be changed.

Metropolis-Hastings Algorithm

The code for this project is located in search/metropolis_hastings. The original code for Metropolis-Hastings is still present but has been abstracted to a specific run configuration for the algorithm, see "How to run" for details.

Original parameters used in the paper by V. van Wieringen:

{
    'add_if_statement_end': 10,
    'add_loop_end': 10,
    'add_token_end': 10,
    'alpha': 1.2,
    'remove_token_random': 20,
    'start_over': 2,
    'type': 'metropolis'
}

To run the alternatives code:
First, make sure that main.py contains [MetropolisHastings, "metro"] in the list of search algorithms. Then, you can either:

1. adjust the code in main.py
2. call main.py with a domain and parameter argument. See "How to run" for an example.

Here is a list of possible parameters:

Parameter	Values	Default	Description
type	"metropolis", "metropolis_hastings", "metropolis_hastings_2"	"metropolis"	Decides which acceptance function to use. Read the paper/code for details
alpha	float >= 0	1	Sets the normalization factor alpha.
add_token_*	natural number	0	Sets weight of the add_token mutation
remove_token_*	natural number	0	Sets weight of the remove_token mutation
add_loop_*	natural number	0	Sets weight of the add_loop mutation
add_if_statement_*	natural number	0	Sets weight of the add_if_statement mutation
start_over	natural number	0	Sets weight of the start_over mutation

Note: Replace the '*' with "end" or "random" to choose the specific locality option