Milestone 2

due at midnight on Sun Feb 19  +125

Compose a cart-pole agent that uses a the Perceptron class to learn how to balance the pole. Instead of using the standard perceptron delta rule, implement a hill-climbing technique, where it randomly tweaks the weights by some small amount, but only keeps the tweaked weights if they perform better than the previously best-performing weights.

There are several parameters to play with:

• By how much can it randomly alter the weights after each trial? (We can think of this as the learning rate.)

• Should it alter all the weights at the same time, or just one (randomly-chosen) weight?

• Should it perform more than one trial (and calculate the average reward) before it adjusts the weights? How many trials?

Try your implementation several times with different parameters. In your repository’s README.md file, document the perceptron’s weights that produced the best performance you were able to achieve, as well as the choices you made with respect to the above parameters.

⇒ Please name your file cartm2.py.

PDF