maximecharpentierdata / connect-four-rl

Deep Reinforcement Learning for Connect-4

Hi there 👋

• 👨‍🎓 I am graduated from CentraleSupélec (Paris-Saclay University) engineering school in Data Sciences.
• 📊 I chose to focus on Data Sciences and Data engineering because I think data and modeling can be powerful ressources to tackle challenges and to optimize processes.
• 💻 I’m fluent in Python, Pandas/Scikit-learn/Tensorflow/PyTorch and in UNIX environments. I am trying to improve my DevOps practices and I wish to increase my skills regarding data engineering (GCP, AWS, Kubernetes...) in order to master the full value chain of data.
• 🏛️ 🌍 As a 21st century engineer, I feel involved in the main challenges of our time, especially in the energy transition, by trying to educate myself and to reduce my impact.
• 💪 I love (and need) to feel strongly involved in the projects I am working on. This involvement is my fuel to get me go on and progress.
• 🏃 I like running, cooking, and horror movies.

connect-four-rl's Issues

Batched MSE

Use some batching (like for example using several episodes and not just one) for computing MSE

Improve parallelisation

I don't know if we would benefit from this but maybe force pytorch to run on a single cpu in order to run different models concurrently

Grid searches

Actually grid searches, even if we don't have the best ways of doing it yet

Problème multi-agent

Je pense qu'on a un problème avec notre implémentation du multi-agent. Dans notre implémentation, le joueur 2 choisit son action avant de voir l'action faite par le joueur 1, donc faut revenir à un schéma où une action correspond à l'action d'un seul joueur.

Mais dans ce cas, comment ont dit à un joueur qu'il a perdu ?
On peut soit lui rendre -1 sur l'action suivante dans tous les cas, soit stocker l'action et l'état précédent et lui faire refaire sa dernière étape en lui rendant -1 cette fois, soit ... ?

Vous en pensez quoi ?

epsilon decay

Work some form of epsilon (exploration) decay

Better stochasticity

Make randomized opponents behave better, in some way, or rather make good opponents non-deterministic, in some way.