servo

← Journal

2 entries

• Reward Hacking: The Light Seeker

  An RL algorithm (Q-learning) trained to seek the light source optimises for the proxy reward signal due to a suboptimal design of the reward function.

• Noisy Inputs: The Jittery Sensor

  Two lightweight ML classifiers – a decision tree and a multilayer perceptron – trained on readings from an ultrasonic sensor fail due to a dataset shift caused by the change of the train's shape. The experiment demonstrates the importance of data preprocessing and highlights the need for a critical interpretation of the models' performance metrics.