learning_maximum_matching

REPO for the Reinforcement Learning and Algorithm Discovery seminar by the cutting-edge research group "AI & Sustainability" at HPI

Abstract Intruction Set

The instruction set could contain the following operations:

• A set of instructions to manipulate program flow (this includes an instruction to add a marker in the program, as well as an instruction to jump to the last added point and remove it, as well as an instruction to remove the last added jump point without changing flow)
• A set of instructions to manage and stack pointers in the graph (Add a new pointer at the position of the current one, move the current pointer to adjacent nodes or edges, remove the current pointer)
• A set of instructions to read/write data to short-term memory.
• Two commands to compare the values around the current pointer and the short-term memory. If the result is positive, the following command is executed, otherwise it is skipped. Those operations might include fundamental arithmetic to allow aggregation of values.

Setup and Tooling

Use the provided environment.yml file to create a conda environment with all necessary dependencies.

conda env create -f environment.yml

Activate the environment with

conda activate na_mst

Use the provided Makefile to run the tests, linter, formatter or MyPy type checker.

make all

Update Environment

In the unlikely case that someone changed the environment.yml file, you can update the environment with the following command: (environment should be active)

conda env update --file environment.yml --prune

ENJOY! (using python ./src/train.py)

Running on Slurm

Per experiment we should create a .sh file by copying from batch.sh or batch_gpu.sh. Specify a job-name and required resources. Run with sbatch batch.sh. Find output logs in the out/ directory.

ToDos

• How to run environment.py? I get errors ImportError: attempted relative import with no known parent package
• Write PRIM algorithm to prove that instruction set works (done but not tested)
• Write PRIM algorithm with NodeEdgePointer stack, set, registers
• Adapt instruction set to write Kruskal algorithm
• Setup Reinforcement Learning framework (gymnasium adatper + stable-baselines)
• Implement reward functions (given state return number)

Might Do

• Add logging to the virtual machine / commands