Tuesday, July 29, 2014

On internal models, consequence engines and Popperian creatures

So. We've been busy in the lab the last few months. Really exciting. Let me explain.

For a couple of years I've been thinking about robots with internal models. Not internal models in the classical control-theory sense, but simulation based models; robots with a simulation of themselves and their environment inside themselves, where that environment could contain other robots or, more generally, dynamic actors. The robot would have, inside itself, a simulation of itself and the other things, including robots, in its environment. It takes a bit of getting your head round. But I'm convinced that this kind of internal model opens up all kinds of possibilities. Robots that can be safe, for instance, in unknown or unpredictable environments. Robots that can be ethical. Robot that are self-aware. And robots with artificial theory of mind.

I'd written and talked about these ideas but, until now, not had a chance to test them with real robots. But, between January and June the swarm robotics group was joined by Christian Blum, a PhD student from the cognitive robotics research group of the Humboldt University of Berlin. I suggested Christian work on an implementation on our e-puck robots and happily he was up for the challenge. And he succeeded. Christian, supported by my post-doc Research Fellow Wenguo, implemented what we call a Consequence Engine, running in real-time, on the e-puck robot.

Here is a block diagram. The idea is that for each possible next action of the robot, it simulates what would happen if the robot were to execute that action for real. This is the loop shown on the left. Then, the consequences of each of those next possible actions are evaluated. Those actions that have 'bad' consequences, for either the robot or other actors in its environment, are then inhibited.

This short summary hides alot of detail. But let me elaborate on two aspects. First, what do I mean by 'bad'? Well it depends on what capability we are trying to give the robot. If we're making a safer robot, 'bad' means 'unsafe'; if we're trying to build an ethical robot, 'bad' would mean something different - think of Asimov's laws of robotics. Or bad might simply mean 'not allowed' if we're building a robot whose behaviours are constrained by standards, like ISO 13482:2014.

Second, notice that the consequence engine is not controlling the robot. Instead it runs in parallel. Acting as a 'governor', it links with the robot controller's action selection mechanism, inhibiting those actions evaluated as somehow bad. Importantly the consequence engine doesn't tell the robot what to do, it tells it what not to do.

Running the open source 2D robot simulator Stage as its internal simulator our consequence engine runs at 2Hz, so every half a second it is able to simulate about 30 next possible actions and their consequences. The simulation budget allows us to simulate ahead around 70cm of motion for each of those next possible actions. In fact Stage is actually running on a laptop, linked to the robot over the fast WiFi LAN. But logically it is inside the robot. What's important here is the proof of principle.

Dan Dennett, in his remarkable book Darwin's Dangerous Idea, describes the Tower of Generate-and-Test; a conceptual model for the evolution of intelligence that has become known as Dennett's Tower.

In a nutshell Dennett's tower is set of conceptual creatures each one of which is successively more capable of reacting to (and hence surviving in) the world through having more sophisticated strategies for 'generating and testing' hypotheses about how to behave. Read chapter 13 of Darwin's Dangerous Idea for the full account, but there are some good précis to be found on the web; here's one. The first three storeys of Dennett's tower, starting on the ground floor, have:
  • Darwinian creatures have only natural selection as the generate and test mechanism, so mutation and selection is the only way that Darwinian creatures can adapt - individuals cannot.
  • Skinnerian creatures can learn but only by literally generating and testing all different possible actions then reinforcing the successful behaviour (which is ok providing you don't get eaten while testing a bad course of action).
  • Popperian creatures have the additional ability to internalise the possible actions so that some (the bad ones) are discarded before they are tried out for real.
Like the Tower of Hanoi each successive storey is smaller (a sub-set) of the storey below, thus all Skinnerian creatures are Darwinian, but only a sub-set of Darwinian creatures are Skinnerian and so on.

Our e-puck robot, with its consequence engine capable of generating and testing next possible actions, is an artificial Popperian Creature: a working model for studying this important kind of intelligence.

In my next blog post, I'll outline some of our experimental results.

Acknowledgements:
I am hugely grateful to Christian Blum who brilliantly implemented the architecture outlined here, and conducted experimental work. Christian was supported by Dr Wenguo Liu, with his deep knowledge of the e-puck, and our experimental infrastructure.

Related blog posts:

Saturday, July 19, 2014

Estimating the energy cost of evolution

Want to create human-equivalent AI? Well, broadly speaking, there are 3 approaches open to you: design it, reverse-engineer it or evolve it. The third of these - artificial evolution - is attractive because it sidesteps the troublesome problem of having to understand how human intelligence works. It's a black box approach: create the initial conditions then let the blind watchmaker of artificial evolution do the heavy lifting. This approach has some traction. For instance David Chalmers, in his philosophical analysis of the technological singularity, writes "if we produce an AI by artificial evolution, it is likely that soon after we will be able to improve the evolutionary algorithm and extend the evolutionary process, leading to AI+". And since we can already produce simple AI by artificial evolution, then all that's needed is to 'improve the evolutionary algorithm'. Hmm. If only it were that straightforward.

About six months ago I asked myself (and anyone else who would listen): ok, but even if we had the right algorithm, what would be the energy cost of artificially evolving human-equivalent AI? My hunch was that the energy cost would be colossal; so great perhaps as to rule out the evolutionary approach altogether. That thinking, and some research, resulted in me submitting a paper to ALIFE 14. Here is the abstract:
This short discussion paper sets out to explore the question: what is the energy cost of evolving complex artificial life? The paper takes an unconventional approach by first estimating the energy cost of natural evolution and, in particular, the species Homo Sapiens Sapiens. The paper argues that such an estimate has value because it forces us to think about the energy costs of co-evolution, and hence the energy costs of evolving complexity. Furthermore, an analysis of the real energy costs of evolving virtual creatures in a virtual environment, leads the paper to suggest an artificial life equivalent of Kleiber's law - relating neural and synaptic complexity (instead of mass) to computational energy cost (instead of real energy consumption). An underlying motivation for this paper is to counter the view that artificial evolution will facilitate the technological singularity, by arguing that the energy costs are likely to be prohibitively high. The paper concludes by arguing that the huge energy cost is not the only problem. In addition we will require a new approach to artificial evolution in which we construct complex scaffolds of co-evolving artificial creatures and ecosystems.
The full proceedings of ALIFE 14 have now been published online, and my paper Estimating the Energy Cost of (Artificial) Evolution can be downloaded here.

And here's a very short (30 second) video introduction on YouTube:


My conclusion? Well I reckon that the computational energy cost of simulating and fitness testing something with an artificial neural and synaptic complexity equivalent to humans could be around 10^14 KJ, or 0.1 EJ. But evolution requires many generations and many individuals per generation, and - as I argue in the paper - many co-evolving artificial species. Also taking account of the fact that many evolutionary runs will fail (to produce smart AI), the whole process would almost certainly need to be re-run from scratch many times over. If multiplying those population sizes, generations, species and re-runs gives us (very optimistically) a factor of 1,000,000 - then the total energy cost would be 100,000 EJ. In 2010 total human energy use was about 539 EJ. So, artificially evolving human-equivalent AI would need the whole human energy generation output for about 200 years.


The full paper reference:

Winfield AFT, Estimating the Energy Cost of (Artificial) Evolution, pp 872-875 in Proceedings of the Fourteenth International Conference on the Synthesis and Simulation of Living Systems, Eds. H Sayama, J Rieffel, S Risi, R Doursat and H Lipson, MIT Press, 2014.

Related blog posts: