# Tensorflow / Deepmind: how do I take actions from observations for math algorithms related to proofs?

Crossposted from here

This question is to ask for directions/suggestions/help on the use of deepmind opensource libraries: DeepMind Lab or TensorFlow in Python.

Consider that I'm new to concepts like deep learning and AI.

Questions are:

• Are there examples on using Deepmind or Tensorflow for math problems where I need to observe values and take actions?

Using an approach similar to to the one described in this page based on observations, actions, rewards, etc., I'd like to call a learing agent to choose among some values. I was thinking to something like this:

• Input: a list of list of tuple (the list will change at each step)
• Action: pick up a value from the input (based on experience)
• Reward: If the value it returned was good or bad for the rest of the algorithm I'm implementing, I'll reward the deep learning agent.

• I cannot train the algorith in advance

The Input is something like this:

edge: (1, 2), face_down: 4, face_up: 5, face_left: 4, face_right: 5
edge: (0, 1), face_down: 4, face_up: 4, face_left: 4, face_right: 5
edge: (5, 4), face_down: 4, face_up: 5, face_left: 4, face_right: 5
edge: (6, 7), face_down: 3, face_up: 5, face_left: 5, face_right: 5
edge: (3, 0), face_down: 4, face_up: 4, face_left: 5, face_right: 4
edge: (4, 1), face_down: 4, face_up: 5, face_left: 5, face_right: 4
edge: (8, 5), face_down: 4, face_up: 5, face_left: 5, face_right: 4
edge: (3, 8), face_down: 4, face_up: 5, face_left: 4, face_right: 5
edge: (2, 3), face_down: 4, face_up: 5, face_left: 5, face_right: 4
edge: (5, 0), face_down: 4, face_up: 4, face_left: 5, face_right: 4
edge: (0, 5), face_down: 4, face_up: 4, face_left: 4, face_right: 5
edge: (1, 0), face_down: 4, face_up: 4, face_left: 5, face_right: 4
edge: (9, 6), face_down: 3, face_up: 5, face_left: 5, face_right: 5
edge: (0, 3), face_down: 4, face_up: 4, face_left: 4, face_right: 5
edge: (7, 9), face_down: 3, face_up: 5, face_left: 5, face_right: 5


The idea is to use the same aproach deepmind uses to play games, but instead of analyzing pixels and use the pad (up, dowm, left, right, fire, jump), is to let the learning agent to analyze some math values and, as the only action, to select one of them.

Are there other approaches or libraries/framework to address such a problem?