Deep Reinforcement Learning Agent (Dra)

Dra is a "deep" reinforcement learning agent that exists in your browser. The reward at every time step is negatively correlated with Dra's distance to the goal. The training network is making predictions of the given reward at each time step given the current Euclidean distance to the goal.

Here are the current parameters:

this.regularizationLambda = 0.005;
this.alpha = 0.001;
this.batchSize = 1;
this.numHidden = 10; // Single hidden layer with 10 neurons

When you open the page, Dra starts updating a fundamental rule set of its environment. When you close the page, it forgets.

Made with paper.js. Inspired by otoro.net.