This repository contains Python implementations of two reinforcement learning (RL) algorithms applied to a grid path finding environment. These implementations were developed as part of a university lecture in machine learning to explore the application of RL techniques in solving grid-based navigation problems.
Consider a robot that starts at a designated starting line within a grid. The primary objective is for the robot to navigate to the target cell as quickly as possible, adhering to certain constraints:
- The robot's movement is constrained to a discrete set of grid positions.
- Velocity is also discrete, representing the number of grid cells moved horizontally and vertically per time step.
- Actions available to the robot are increments to the velocity components, allowing for movement in all directions.
- Each velocity component can be changed by +1, -1, or 0 in each step, resulting in a total of nine (3 x 3) possible actions.
- Both velocity components are restricted to be less than 3, and they cannot both be zero except at the starting line.
- Each episode begins with the robot positioned randomly along the starting line, with both velocity components initialized to zero.
- The episode terminates when the robot successfully reaches the target cell.
- The robot receives a reward of -1 for each step taken until reaching the target.
- If the robot collides with obstacles or walls, it is reset to a random position along the starting line, both velocity components are reduced to zero, and the episode continues.
- The episode also ends if the robot reaches the target cell.
- First Visit Monte Carlo Control - Tabular algorithm
- Deep Q-Learner - Temporal Difference Off-Policy algorithm with neural network as approximator
Grid problems are loaded using .txt files in the grids folder. To run the a learning algorithm with a grid, set env_text = "grid_simple" to the filename.
Policies and training progress of already trained grids can be found in the plots folder.
