Showing posts with label reward function. Show all posts
Showing posts with label reward function. Show all posts

Sunday, October 29, 2023

The Road to DeepRacer Victory: Winning Strategy Insights

The Starting Line – Embracing the Challenge:

When our company announced the DeepRacer competition, a rush of excitement mixed with a hint of nervousness surged through me. Despite my unfamiliarity with DeepRacer, the adrenaline of competition and the allure of uncharted territory beckoned me.

The First Lap – Understanding the Basics:

Before diving in, it was imperative to grasp the foundational concepts of DeepRacer. Reward function, hyperparameters, and action space settings were pivotal. The learning curve was steep but worth every effort.

Mapping the Route – Reward Function Strategy:

The chosen reward function for our DeepRacer prioritized track adherence, efficient steering, and speed optimization based on steering angles. Corner-cutting was not just allowed but encouraged, aiming for both safety and aggressive driving. Behind this strategy, there were auxiliary functions that ensured smooth, efficient navigation with a keen focus on certain track points.

Tuning and Iteration:

Creating the model was just the start. Running it on a virtual track, analyzing the logs, and then refining the approach was an iterative process. Using the Chatgpt Advanced Data Analytics plugin and DeepRacer analysis and garnering insights from this platform streamlined the tuning process. Iteration after iteration, tweaks to the hyperparameters and action space led us closer to our goal to get a minimum of 9sec.

Achieving Top Speeds:

The pinnacle of our efforts saw us achieving a time of 8.59 seconds on the virtual track, placing us at the top among our company competitors. The euphoria was short-lived, however, as we were surpassed by mere milliseconds the following morning.

Game Day Showdown at AWS re:Invent 2018:

The D-day was nothing short of spectacular. A key revelation was the necessity for manual speed override, allowing the model to focus on steering. Our very first run clocked an impressive 8.17seconds, outpacing our virtual best. Yet, the competition was fierce, with the fastest run of the day being 7.69 seconds.

In conclusion, the DeepRacer journey was an incredible learning experience. From conceptualizing and refining our model to facing unexpected challenges on Game Day, each step brought its own set of lessons. The world of AI and machine learning is ever-evolving, and our journey with DeepRacer served as a powerful reminder that innovation, perseverance, and adaptability are key to success.

For a firsthand look at the electrifying race, watch my race video here

For enthusiasts and fellow racers wishing to delve deeper, you can access the analytics tools I used here

and find my code on GitHub

Optimizing My Model: The primary takeaway is that all lines should exhibit a steady upward trajectory, with the red line maintaining as close to 100% as feasible.


Follow me on 


Navigating Ethical Waters: A Day in the Digital Life of LLM's

Introduction Greetings from your AI companion, GPT-4! Today, I'm taking you behind the scenes of my daily routine, which has recently be...