What is the AI discipline where an algorithm learns from an initial training set, but then refines its learning as it uses that training?

Question

Imagine a system that is trained to manipulate dampers to manage air flow. The training data includes damper state and flow characteristics through a complex system of ducts. The system is then given an objective (e.g. maintain even flow to all outputs) and set loose to manage the dampers. As it performs those functions there are anomalies in the results which the system is able to detect. The algorithm CONTINUES to learn from its own empirical data, the result of implemented damper configurations, and refines its algorithm to improve performance seeking the optimum goal of perfectly even flow at all outputs.

What is that kind of learning or AI system called?

Answer 1

I believe this can best be done with reinforcement learning via Deep Q Learning. That's where I would start. Steps are:

• Initialize a Q table.

• Choose an action.

• Perform the action.

• Measure the reward.

• Update the Q.

A neural net will approximate the Q function. See: https://medium.com/emergent-future/simple-reinforcement-learning-with-tensorflow-part-0-q-learning-with-tables-and-neural-networks-d195264329d0

Also consider policy gradients, actor critic, and inverse reinforcement learning.