Pie & AI: The AI Tournament - Learn to train RL Agents for Video Games

Learn to train RL agents for video games and win 1000 CHF (first prize)!
In this event we will discuss all you need to know to get started.

This is a free ONLINE event, join us from wherever you are in the world.

Agenda
18:00 Sign in to the webinar
18:05 Alessandro Palmas (Artificial Twin): "The AI Tournament & How to train RL Agents for Video Games"
18:45 Q&A
19:00 End of zoom session

How to train RL Agents for Video Games:
Video games are the perfect setting for Reinforcement Learning application and research: they are virtual environments by definition and, as such, they allow for optimizations like parallel training and a 100% match between training environment and real world. They also have an additional bonus: they are super fun!
As usual the key starting point is to formalize them as Markov-Decision-Processes, clearly identifying observation and action spaces. Having done that, most advanced techniques can be applied. Especially on most recent games, it is needed to leverage deep reinforcement learning, given the high dimensionality of the input.
Depending on the specific nature of the observations and action spaces (e.g. pixels or numbers, discrete or continuous) different deep learning architectures may be needed, and additional advanced techniques, such as Imitation Learning or self play, can play an important role.
It is this huge passion for reinforcement learning and video games that gave birth to DIAMBRA | Dueling AI Arena (diambra.artificialtwin.com).

The AI Tournament:
DIAMBRA and RLZ just launched the AI Tournament, where the challenge is to train a RL agent to effectively play Dead Or Alive++. Agents will be evaluated once a week during the two months competition period. The best five will qualify for the final event in which they will compete live for the final prize.
We will discuss all technical details: the observation space configuration and how to use it, the four possible action spaces available and the default reward function. We will take a look at the tournament repository, see how to interact with the environment and how to compose a valid submission.
Finally, we will compare a random agent with a trained Proximal Policy Optimization one, to demonstrate environment learnability.

About Alessandro Palmas:
Enthusiastic aerospace engineer, with a great passion for artificial intelligence and machine learning. 9+ years of expertise in software development for scientific applications and complex software systems.
Co-Founder of innovative initiatives, including one of the first cloud-based platforms to perform fluid dynamics and stress analysis simulations within a web-browser, and a customized service for machine learning, computational geometry and physical modeling applications.
Addicted to the Reinforcement Learning world, in particular to its applications in video games; Co-Author of Packt’s Book “The Reinforcement Learning Workshop”, and creator of DIAMBRA, a place to watch RL Agents being trained to play video games and competing against each other.

This is a Pie & AI event in partnership with DeepLearning.AI.
Pie & AI is a series of DeepLearning.AI meetups independently hosted by our global AI community. In addition to signing up on Meetup.com, sign up here (https://www.eventbrite.com/e/150600587411) if you'd like to receive a course promo code and future event updates from DeepLearning.AI

Organizers
Claus Horn (ZHAW), Georg Russ (die Mobiliar), Mark Rowan (Rowan Cognitive Data Science Solutions)

Location
ONLINE event

Sponsors
rockstar* recruiting
Swiss Re Institute