For my final project for Peter Stone's Autonomous Robots class, I'm going to be developing an improved goalie behavior. To accomplish this, instead of hand-tuning velocity thresholds, I'm treating shot detection as a classification problem. Additionally I'm looking to remove more and more of the noisy features we currently use (e.g. velocity, position, and the Kalman filter in general). View the wiki on the right hand side for more details.
As a side note, I will be implementing all of the models I use from scratch. Here are the ones I have planned:
- Linear Regression
- Softmax
- Fully-Connected Neural Network
- Recurrent Neural Network
I'm implementing linear regression and a fully-connected neural network because they're building blocks to implement softmax and RNNs, respectively.
For each classifier, in order to gain intution, I'll implement a 1d version of it (no vectors nor matrics) and then extend it to handle arbitrarily-sized input. I was heavily inspired to take this approach by Andrev Karpathy's Backpropagation tutorial. Thanks Andrev!