After competing in multiple challenges on Codingame with bots using simple heuristics, it's time for something new: the goal is to learn and advance to the Gold league in an upcoming challenge.

git clone https://github.com/tonigineer/nn-codingame.git ~/journey && cd ~/journey

# Optional, but recommended
python -m venv .venv
source .venv/bin/activate

pip install -r requirements.txt

Using the MNIST dataset to train a simple Neural Network with TensorFlow and then transferring the parameters to a Neural Network implemented from scratch.

python feasibility/example_ann_tf.py

...
Epoch 10/10
1875/1875 ━━━━━━━━━━━━━━━━━━━━ 1s 531us/step - accuracy: 0.9960 - loss: 0.0140
313/313 - 0s - 459us/step - accuracy: 0.9794 - loss: 0.0787

The weights and biases have been exported. The ANN, implemented in example_ann_scratch.py, also includes backpropagation and gradient descent, which are not necessary for this task.

python feasibility/ann_with_tf_param.py

...
Accuracy with test data: 97.94%

✔️ Both evaluations yield the same accuracy on the test data from mnist = tf.keras.datasets.mnist. Therefore, transferring weights and biases to a custom implementation from scratch proves to be effective.

Lessons learned for the next parts:

• Adapt sizes to TensorFlow conventions from the outset
• Implement a dynamic architecture based on weights and biases

Todo

• 100k limit on characters from Codingame
• plain text export of weights to use on Codingame
• Reduce size of numbers