added more training details

README.md

@@ -30,6 +30,15 @@ The RL Zoo is a training framework for Stable Baselines3
 reinforcement learning agents,
 with hyperparameter optimization and pre-trained agents included.
 
+## Training details
+
+This might seem like cheating but ... 
+it doesn't feel right to burn 90 minutes of GPU time (and tying up a colab machine) to reporduce a model that is already available on the hub ... 
+so I've borrowed a trained model and continuted training for just a few steps.
+
+I'm wondering if dqn training could be made faster/more efficient if we train the CNN in a semi-supervised style rather than starting with random weights ... 
+i.e given the previous 4 frames, can the model predict the next frame ... 
+
 ## Usage (with SB3 RL Zoo)
 
 RL Zoo: https://github.com/DLR-RM/rl-baselines3-zoo<br/>