quantuminformation / youtube-space-invaders

Follows along the youtube video series @ https://www.youtube.com/playlist?list=PLCrwuqjmVebIU2b4qesWz7PAEm1UJl15n

youtube-space-invaders

Simple space invaders game for the browser, getting more fun each release!

You can play the latest version of the game here

I used TypeScript as it is pretty nice for games and catches plenty bugs at compile time.

First video of the series

Latest video of the series

Conventions

Angles are measured clockwise from north. For example, east is 90 degrees.

How to use in your own website

I published this over npm. All you need is to give the game object a 2D Canvas object and it will be added to your website.

youtube-space-invaders's Issues

Add option to disable TensorFlow.js reinforcement learning agent (bot)

Add testing data for actor and critic networks

Add controllable limit to size of experience replay buffer

Training slows down over time as more state/action pairs are added.

Add random noise to agent action outputs to allow exploration of possible action space

For more information, see the Deep Deterministic Policy Gradient section of the following Medium article: https://medium.com/@jonathan_hui/rl-reinforcement-learning-algorithms-quick-overview-6bf69736694d

Agent API for tensorflow.js agent

Add speed control in agent interpreter panel

Add reward/loss over time graphs for actor and critic networks

The actor reward graph should display both the predicted loss generated by the critic network (equivalent to the actor optimization loss) and the actual loss once the training episode is complete.

Embed on websites

Would be cool to allow people to embed this game on their sites, eg in a 404 page.

Add button to reset game

Mobile-friendly

I need to provide touch controls.

Add visualization of convolutional layer outputs

Use maxPooling2d instead of avgPooling2d

The issue preventing me from using tf.layers.maxPooling2d has been resolved: tensorflow/tfjs#1189.

End game when ships reach bases

When in bot mode, also reset game when player wins

Add learning rate control

Add actor network training system

Optimize the parameters of the actor model to maximize the predicted reward value generated by the critic.

Add pre-trained agent models