Machine Learning
In 2019 I had the good fortune to be able to use some company hack days to see about learning how various flavours of reinforcement learning work.
In the end I came away disillusioned, having found that there’s very little insight available into how a lot of the systems work or how to diagnose them when they don’t. Most of the guidance was about “changing parameters until you get better results”, but with slow cycle times and strange results after running algos for longer periods I decided to put this avenue on ice and stick to writing game AI myself.
Spent some time over the weekend having a go at implementing a Q-learning neural network. Limited results, but certainly felt like I gained a better foundational understanding of #DeepLearning and #NeuralNetworks. #madeWithUnity pic.twitter.com/eORTgVuphV
— Jez 🌻 (@NinjaJez) March 25, 2019
Cracked the bugs with my first solo #deepLearning #neuralNetwork! Discovered I was getting double value overflows in my activation function, so put in sensible rounding thresholds. With the algo working I focused on reward functions and agent selection and finally got results! :D pic.twitter.com/41vXA50dVW
— Jez 🌻 (@NinjaJez) April 1, 2019
Attempting to extend my #NeuralNetworks to handle 3D; much slower going than with my 2D setup, so might need to play around with a stronger reward function, better fitness selection, different state params or a different network topology. #parameterPoker #machineLearning pic.twitter.com/sdiJKg4etO
— Jez 🌻 (@NinjaJez) April 2, 2019
On a hunch I increased the lifespan of the agents and on the next run saw a massive improvement in training success! Yay!
— Jez 🌻 (@NinjaJez) April 2, 2019
Still seeing a lot of agents seemingly intentionally turning the wrong way, which bears investigating :/#MachineLearning #deepLearning #unity3d pic.twitter.com/QXbnWSOYeg
Trying to work out how to debug a #neuralNetwork, even one as simple as mine, is a pain. Not sure why my agents are so scribbly: possibly cos the angle states I'm giving them aren't correctly correcting for the agent's rotation? Frustratingly opaque!#madeWithUnity #deepLearning pic.twitter.com/MuJRBhCRiV
— Jez 🌻 (@NinjaJez) April 3, 2019
Managed to get myself up and running with #unity3d #mlAgents! Certainly yielding better results than my little DIY network haha! #machineLearning pic.twitter.com/BglF6UwBhd
— Jez 🌻 (@NinjaJez) April 25, 2019