Interacting with Grid World: Implementing a Random Agent | CodeSignal Learn

Interacting with Grid World: Implementing a Random Agent | CodeSignal Learn

Introduction

Welcome to final lesson of our course "Environment Engineering: The Foundation of RL Systems"! In our journey so far, we've built a solid foundation by understanding the core concepts of Reinforcement Learning, and we've implemented a complete Grid World environment with__init__, reset, step, and render methods. Now that we have a fully functional environment, it's time to put it to use! In this lesson, we'll learn how to create a simple random agent that will interact with our environment . We'll run multiple episodes, track progress, and visualize results — bringing our Grid World to life!

Random Agents as Baselines in Reinforcement Learning

First, let's understand what a random agent is and why it's an important starting point for RL projects. A random agent is exactly what it sounds like: an agent that selects actions randomly from the available action space, without considering the current state or any learned strategy. So, why implement a random agent when our goal is to build intelligent systems? Baseline Performance : Random agents provide a minimum performance benchmark. Any learning algorithm should perform better than a random agent by definition. Environment Testing : They're excellent for validating that our environment works correctly before implementing complex algorithms. Exploration Properties : Random agents naturally explore the entire state space given enough time, helping us understand the dynamics of our environment. Simplicity : They require no training or complex decision-making process, making them perfect first agents. In our Grid World, a random agent will wander aimlessly, occasionally finding the goal by chance. This process will help us demonstrate the complete agent-environment interaction loop that forms the foundation of all RL systems.

Defining the Main Function

Let's start by creating a main function that will set up our environment and define the basic structure for running episodes: Pythondef main(): # Create our environment with a 5x5 grid env = GridWorldEnv(size=5) # Define possible actions the agent can take actions = [0, 1, 2, 3] # up, down, left, right # Create human-readable action labels for better output action_meanings = {0: "up", 1: "down", 2: "left", 3: "right"} # Set how many complete episodes we'll run num_episodes = 3 This code creates an instance of our GridWorldEnv with a 5×5 grid and defines all available actions and respective mappings for our agent as well as the num_episodes our main loop will run for.

Sign up

Join the 1M+ learners on CodeSignal

Be a part of our community of 1M+ users who develop and demonstrate their skills on CodeSignal

Implementing an Episode Loop

Now that we have our setup, let's implement the actual loop: Pythondef main(): # Setup code from previous section... for episode in range(num_episodes): # Reset environment to get initial state state = env.reset() done = False total_reward = 0 print(f"\n=== Episode {episode+1} ===") print("Initial State:", state) env.render() This sets up each episode with a clean slate. Recall that the reset() method returns us to the starting position (the top-left corner at (0,0) in our Grid World), which allows the agent to begin a fresh attempt at reaching the goal.

Creating a Random Agent

Now, let's implement the random agent and its interaction with the environment: Python while not done: # Select a random action action = random.choice(actions) # Take the action in the environment next_state, reward, done, info = env.step(action) # Update our cumulative reward total_reward += reward This code continues taking steps until the episode is done (either by reaching the goal or by exceeding max steps). Our random agent simply picks an action randomly with no strategy with random.choice(actions), which is then executed via env.step(action) . We also update the cumulative reward after each agent-environment interaction.

Tracking Episode Progress

Finally, let's enhance our code to display what's happening at each step and summarize the episode results: Python # Print detailed information about the step print(f"Action: {action} ({action_meanings[action]}), Next State: {next_state}, Reward: {reward}, Done: {done}") # Check if we hit the max steps limit if info.get("timeout", False): print("Episode timed out!") # Visualize the current state env.render() # Display a summary of the completed episode print(f"Episode finished in {env.steps_count} steps with total reward {total_reward}.") This code provides both real-time details as the agent moves through the environment and a concise summary when the episode concludes. The step-by-step output helps us track the agent's journey, while the episode summary (total steps and reward) gives us metrics to evaluate performance. Without this visualization and summary data, it would be difficult to understand why the agent succeeds or fails in different episodes.