Alexander Whillas

Free Energy Principle

Started ; Updated Aug 10, 2023

The basics

Free Energy Principle, Basic outline

So the Free Energy Principle (FEP) is a sort of universals objective function for cognitive agents i.e. agents with some sort of cognition or internal world representation that learns. It learns from sensory input and some actions that it can enact on the world.

Sensation -> Internal model -> Actions -> World changes -> Sensations -> ... etc

Its a feedback loop. What they are calling free energy is the difference between the predicted (by the model) sensor(s) state and the actual sensor(s) state. The system tires to minimise this difference, or “free energy”, or entropy, or “surprise” as its call in relation to the FEP by ether updating its model or by using actions on the world, or both.


How does it decide which? How do you give it an objective like walk, do the laundry or make money on the stock market?

I feel like this feedback loop should give the agent the greatest self control possible, but something else is missing to give it higher order objectives.

My RoboCup Experience

Robocup Standard Platform league

So I was vaguely involved in the RoboCup1 at UNSW when I was doing my masters back in 2013. The control systems on the bots then was very crude. I thought that UNSW would be using some sort of reinforcement learning (RL) to control walking but they were actually hand programming every movement.

I realised there where all these sensors on the “Standard Platform” bot that we were not using and my thought at the time was that if we feed these into a feedback loop into a learning algorithm like RL2 with the objective of keeping itself upright then it might be able to learn to walk. Further more, since the robots were getting damaged, quite easily from falling over a lot, individual robots might learn how to deal with their own particular damage configuration without any addition programming.

Does this lead to consciousness?

This leads to pondering if this is a system for understanding how consciousness might emerge i.e. if the system then has a model of itself in the world it is modelling, in fact, it would be hard to imagine how it would operate without it.

Feedback is the key

From my days of making experimental music/noise I know that feedback is very powerful. It was what always made the most interesting results. Here too I think the feedback mechanism is a key factoring in making this work. The system balances itself.

  1. Competition between universities to make robots play soccer. I’d really like to get my hands on a simple robot I could program and try some of these ideas out on! 

  2. I would obviously say a deep neural network now.