RL-Sandbox

Experimentation bed to get the team familiarized with RL

Gridworld environment

This repository contains an implementation of a simple gridworld environment to test if an RL algorithm is working correctly. The environment consists of a (n,n) grid. The objective is to reach the destination square, which can either be set to a fixed position or randomized. The observation is the (x,y) coordinates of the current agent position and of destination square (4,) vector. There are 4 possible actions in the action space: up, down, left, and right. The agent receives a -1 reward at each time step, regardless of the action. The agent always starts in the (0,0) square in the grid.

The gridworld class implements two functions: reset() and step()

• __init__(grid_dim, randomize_goal, goal_position) - Initializes the RL environment

  • Arguments:
    • grid_dim - Dimension of the grid world, default: (10,10)
    • randomize_goal - Whether or not to randomize the goal square when the environment is reset, default: False
    • goal_position - If the goal is not randomized, the fixed position of the goal, default: (9,9)

• reset() - Resets the RL environment to start a new simulation

  • Return
    • observation - The initial observation for the RL environment

• step(action) - Moves the RL enviroment forward 1 time step:

  • Arguments:
    • action - The action selected by agent
  • Return
    • observation - The observation for the next time step
    • reward - The reward received for the action
    • done - Boolean describing whether the agent has completed the task
    • info - Contains stats about the simulation
      • Total number of steps
      • Total reward