I find the python code very clear, but I would prefer to see a real life interesting application that doesn't require a lot of computation. In a post in wildml there is an example of using NLP and deep learning for a simple task but after 22 hours of computation the final result is a little disappointing to say the least.
I like to read wildml.com and fastml.com blogs, but I would like to find more simple applications that shows real value without using lots of resources. Perhaps there is a subfield of RL where using some kind of proper human intelligence one can hope to beat those giants provided of unlimited computational and financial resources
I gave a talk a PyConSG this year[1], which included a demonstration of training a Reinforcement Learning model on a 'Bubble Breaker' game. There's also more detail available[2].
The Jupyter notebook is included in the GitHub repo[3], and includes a 'scaled down version' that takes ~5mins to train on a MacBook's CPU. There's also a downloadable 'full scale' model that was trained in ~7hours on a Titan X. It plays the game (on average) better than me...
Most RL algorithms are polynomial time or worse, and they use large datasets. Computation is always going to be an issue which is why most successful implementations are around simplifying models and datasets.
If you can figure out a way of making RL better than polynomial time there is at least a Turing Award for you.
I like to read wildml.com and fastml.com blogs, but I would like to find more simple applications that shows real value without using lots of resources. Perhaps there is a subfield of RL where using some kind of proper human intelligence one can hope to beat those giants provided of unlimited computational and financial resources