Artificial Intelligence for Robotics

Adaptive learning rate

In our first example, we used a constant learning rate of three units to move our robot arm. This equates to about 2.1 degrees of movement from the motor. This step is the smallest and largest single step motion we took at each part of the learning process. What if we changed the learning rate adaptively? As we get closer to the goal, we can make the steps smaller, and if we get farther away, we can take bigger steps. We insert a line at each update just after where we calculate the reward, as shown here:

state,reward = robotArm.step(action,learningRate)# insert this new line to try an adaptive learning ratelearningRate = (100-reward)/2.5

The maximum reward for reaching the goal is 100 points. We want the learning rate ...