Amazon Cloud Technology China Summit: Deep Learning Amazon DeepRacer

preamble

What is Amazon DeepRacer?

Amazon DeepRacer is an autonomous driving simulation racing platform based on deep learning and reinforcement learning technology launched by Amazon. It provides a cloud simulation environment and a physical racing model, allowing users to control the driving of the racing car by writing code and training the model, so as to learn and apply artificial intelligence technologies such as deep learning and reinforcement learning.

The Amazon DeepRacer platform includes the following main components:

1. Racing model: The DeepRacer racing car model uses a 1/18 scale model body, equipped with a series of sensors, cameras, motors and other hardware devices, and can be connected to the cloud simulation environment or local development environment through a wireless network.

2. Cloud simulation environment: DeepRacer provides a simulation environment based on Amazon Cloud Services. Users can develop, train and test models in the cloud without purchasing expensive hardware and building a local environment. The simulated environment includes multiple tracks, different weather and lighting conditions, different car setups, and more.

3. Development tools: DeepRacer provides a variety of development tools and frameworks, including Python, Amazon machine learning services, etc., to facilitate model development, training and deployment for users. Users can use these tools to write code, tune hyperparameters, modify reward functions, and more.

4. Leagues and communities: DeepRacer also provides DeepRacer leagues and community events around the world, allowing users to compete and communicate with other developers. Users can submit their own models, participate in competitions and challenges, and earn rewards and recognition.

Amazon DeepRacer aims to let more people understand and apply artificial intelligence technology, and promote technological innovation and community cooperation. It can help developers quickly master deep learning and reinforcement learning technologies, and explore applications in areas such as autonomous driving and intelligent control.

Example of Reinforcement Learning Model Construction

Reinforcement Learning Process for Amazon DeepRacer

In the console of Amazon DeepRacer, you can see that there are various virtual track environments and virtual racing cars online. What the user needs to do is to set the action for the car, set the reward mechanism, and then let the car run on the track over and over again for self-learning and intensive training. After many times of training and parameter optimization, an optimal model is selected to compete with other players.

As early as 2021, I participated in a complete online competition, so I can tell you responsibly that getting started is not difficult. Let me share with you my experience and experience of participating in Amazon DeepRacer:

As a novice for the first time, I first learned about the introductory guidance and rules through the Amazon DeepRacer homepage, and then tried to train the first model.

After the simulation evaluation on the simulated track, it was found that the wobbly car could run a complete lap, but it ran outside the circle in the next two laps.

So later, the training time was increased, the parameters and reward rules were slightly modified, and a new model was retrained.

Submit the model for the first competition with other contestants, out of 65 contestants, my rank is 40. It's not the last place, I feel very happy and surprised.

The practical process of Amazon DeepRacer

The specific operation process of reinforcement learning model construction is as follows:

create model

Open the Amazon DeepRacer console, and after understanding the basic principles, click the Create model button to start creating the model.
When creating a model, first confirm to set a model name and environment simulation information.
The Environment simulation section selects the track for training. For the June 2021 competition, our chosen track is "re:Invent 2018".

Configuration Environment

Select the race type of the model (Race type), set the training algorithm and hyperparameter and select the race vehicle.

Among them, there are three types of competition modes: pure racing, obstacle avoidance racing, and vehicle avoidance racing. We choose the simplest pure racing mode by default.
The default settings can be used for the training algorithm and hyperparameters, and you need to try different parameter combinations if you want to improve the performance later. Since the learning framework is Tensorflow, it will involve related algorithms and hyperparameters in the Tensorflow framework, and the default algorithm is PPO. Other hyperparameters include: gradient descent Batch size, learning rate, loss function, etc. For these hyperparameters, default values ​​can also be used in the first experience.
Regarding the racing vehicle, there is a racing car in your garage by default, and you can also define your own car, including choosing the color of the racing car and the steering angle of the action.

write function

The reward function is the key to the model. The reward function is essentially a combination of training racing rules, which is related to the quality of the final trained model.

Programming experts can even calculate the optimal path of the track, adjust the speed based on the optimal path and give corresponding rewards, so as to train and obtain a more accurate model.

Friends who are not familiar with programming languages: there will be a simple reward function by default, which can be used directly without modification.

Then you can click the "Create model" button to create your first model.
After the model is successfully created, it will go through a few minutes of initialization process, and the system background will automatically prepare resources and start training according to the settings of the created model.

After starting the training, you can observe the training process in real time through the console, as shown in the figure below, the left side of the picture shows the real-time completion of the model training, and the right side shows the video of the real-time movement trajectory of the car on the track.

model evaluation

After the model is trained, we can evaluate the effect of the model training through the console, and click the Start evaluation button to start the model evaluation.
After the evaluation is completed, the results are shown in the figure below. If the Trail results are 100% and the Status is Lap complete, you can use this model to compete with other teams.
After a complete experience of the training and competition process, the remaining work is to improve the quality of the model. By constantly trying different combinations of reward functions, algorithms, hyperparameters, and racing parameters, continuous optimization is performed until the optimal model is trained.

2023 Amazon Cloud Technology China Summit

The 2023 Amazon Cloud Technology China Summit is an annual large-scale technical event held by Amazon Cloud in China, hoping to provide a platform for practitioners and users in the cloud computing industry to communicate and share.

Special Events

From June 27th to 28th, the finals of the Amazon DeepRacer China Summit will be held, and the Amazon DeepRacer Autopilot Famous School Invitational Competition will be launched simultaneously. Amazon DeepRacer training camps, on-site challenges and Girls in Tech Show will also be planned on site. At the same time, the Global League is officially open, and everyone is welcome to sign up for free.

• China Summit Finals: On June 27, companies from different industries will hold the China Summit Finals.
• Amazon DeepRacer Autopilot Famous School Invitational Tournament: On June 28, it will collaborate with colleges and universities to play a top-level artificial intelligence technology competition specially designed for Chinese college students, looking for future "new forces".
• Amazon DeepRacer Training Camp: Build your first reinforcement learning model on the spot and quickly master machine learning knowledge.
• Girls in Tech Show: Witness the excellence of female racers.
• Global League: This year's Global League has officially started. The top drivers from all over the world have been selected and promoted through layers of selection, and finally competed for the top in the Amazon cloud technology re:Invent conference. In addition to generous prizes, the monthly champion of the Greater China Division will go directly to Las Vegas every month, and the travel expenses for the direct travel will be fully sponsored by Amazon Cloud Technology! The best score at the China Summit will be directly promoted!
  
  In addition, there are many other exciting activities waiting for you to go to the scene to unlock.

Registration address

Hurry up and click the link below to sign up for the 2023 Amazon Cloud Technology China Summit for free, and participate in the event on-site to receive various official gifts!

1. PC terminal link: click to register
2. Mobile terminal link: Click to register