Skip to content

Open-source code repository where reinforcement learning agents solve grid-maze environment and represent learning process as a graph learning.

Notifications You must be signed in to change notification settings

wsojka00/graph_navigation

Repository files navigation

Learning representation of the structured environments through a complex graphs inference

Bachelor Thesis
Cognitive Science
Faculty of Philosophy and Social Sciences
Nicolaus Copernicus University


Abstract

Humans and animals succeed in searching through various complex environments. Back in the thirties, Tolman argued that such abilities arise from the construction of “cognitive maps” through exploration. Such maps enable active mapping of real-world environments to simplified cognitive representations. Although, this theory is widely accepted and studied, the way of how different cognitive maps are stored in the brain is still discussed. Some authors proposed that cognitive maps represent the environment as Euclidean space, but in my thesis I provide evidences that support the hypothesis that graph-theoretical approach provide a more robust framework that additionally allows for behavior flexibility. Here, I investigate a potential of reinforcement learning algorithms trained in navigation of different environments to construct an internal graph-based representation of these spaces. An ability to form a cognitive map in parallel with real-time exploration may underlie human ability to navigate in various spaces, represent the world structure and generalize to novel tasks.

Folders
  • algorithms with possible solutions that can be applied to a grid-maze environment
  • mazes configurations of a 2D grid stored in .txt file
  • simulations ran on given maze together with the simulations recordinds
  • tests consist of code used to practice different libraries
  • thesis
Pipeline

01_create_maze: how to create various mazes using numpy or loaded mazes from txt files. See examples in folder graph_navigation/mazes/src.

02_algorithm_run: apply defined solvers to various mazes. See examples in folder graph_navigation/algorithms.

Resources

Big thanks to Xingdong Zuo for creating mazelab framework that speeds up the process of starting with the task of putting an agent in the maze in which it has to navigate. Besides, I would like to strongly recommend acknowledge all interested people to check neuronav created for the purpose of standardization of spatial navigation tasks in various environments using multiple RL algorithms from recent papers.

About

Open-source code repository where reinforcement learning agents solve grid-maze environment and represent learning process as a graph learning.

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published