AutoML as RL game

Branch 'pretrainedCNN_with_extended_actions'

Using pretrained CNN and extra action 'done' to let the agent decide when to switch images

The architecture:

Extended action space:

We added binary choices 'done' as part of the actions that allows the agent to decide when to terminate an episode, get its classification checked and switch to the next. The extended action space = {Up, Down, Left, Right, done}. done is 0 by default, the agent set it to 1 when it wants to terminate.

Results:

Generated by running test_pretrained_CNN.py

BATCH_SIZE = 128

NUM_STEPS = 10

GAMMA = 1 - (1 / NUM_STEPS) # Set to horizon of max episode length

EPS = 0.05

NUM_LABELS = 2

WINDOW_SIZE = 8

NUM_EPISODES = 5000
TARGET_UPDATE = 10
RUNS = 3