Hacker News Reader

Comment by technotony | Hacker News Reader

This is fascinating. It proposes that intelligence is about maximizing the control we have over future events, ie maximize the entropy of the system. Intuitively this aligns well with how I think about many strategic options or inflections in life/business - you try to take the path which maximizes future options. Lots of interesting applications, will be interested to see how well they manage to communicate this work across different disciplines.

View on HN

I think it's about trying to minimize entropy (disorder) of the system. Maximum entropy is maximum disorder (which the universe tends toward naturally), such as when the pendulum hangs below the cart without any intervention.

Why is this down voted? It is correct. The article worded it wrong. A system with lower entropy has more future states than a system with high entropy.

Yeah, I had to parse the article a couple times because the wording is wrong. It seems to me that entropica is trying to minimize the entropy of the system. Life is all about low entropy.

Indeed. I'd say intelligence tries to keep its environment in a state it can predict.

I think the trick to understanding this is that, in order for an intelligent being to maintain it's internal homeostasis, it must (in an entropic sense) really screw up it's surrounding environment. Thus, if you look at the whole system, the intelligent being is maximizing entropy. Phrased another way: Because it wants to maximize it's own homeostasis, it has to work extra hard to achieve this. Work creates entropy when looking outside the whole system, and bam - you have entropy maximization.

I'm not sure I'm right, but this logic works for me...

Research just observes that if you try to keep your environment in the state in which you have maximum number of possible futures you can choose from by acting in specific way, then your behavior leading to this state seems intelligent.

Good one: Ralf Der and colleagues proposed homeokinesis as such a principle in 1999. That turned out not to be enough, because the agent would then have an incentive to move into a corner and just hide. That renders the world very predictable, but clearly not a good strategy on the long run.

So, they generalized the method to find a predictable future, but in very unstable states. Later, Der and Ay generalized this towards maximization of predictive information - for that, it is not sufficient to predict to environment, but there also must be something nontrivial to predict. This gives rise to a number of highly interesting behaviours in many of their scenarios.

So, yes, your idea is a good one...

Well, the abstract of the actual paper (linked up-thread) uses the phrase "entropy maximization". I still don't understand entropy after taking a physics class on it, but I did gather it's more complicated than "order" vs "disorder" in any intuitive sense.

It is also seems compatible with some experimental parts of economy, such as the study of soldiers of different countries trading MRE which suggest there is a value in diversity, not just taste or quantity.

If there wasn't, why would say french soldiers be exchanging their wine and pate de foie gras MREs - especially after having satisfied their curiosity of what other MREs taste like, ie after the first trade?

Your analysis is not fully accurate, if I understand the paper correctly; I believe you do not maximize the actual8 entropy of the system. You maximize the potential* future entropy of the system. That's something entirely different.

There is actually a line of research where channel capacity of the actuation/perception channel is being used to model that (it's also named "empowerment", in analogy to the social science term, but it is measured in terms of information theoretical quantities). Here, one does not only model the entropy of the future, but in fact that amount of entropy about the future that the agent can systematically control. Or, more precisely, how well the agent is able to systematically control its future (which allows it maximal future entropy - but only entropy it can actually itself generate!)

See e.g. Klyubin et al. 2005, 2008, Capdepuy et al. 2007 and later, Jung et al. 2011, Salge et al. 2012. Indeed, maximizing the control over potential futures (empowerment) seems to work well for a range of scenarios, including the discovery of points of high centrality in mazes and graphs, as well as pole-balancing, acrobot balancing, bicycle balancing examples, in short, survival scenarios; also for discovering new object manipulation modes, driving sensor evolution models, self-organizing multiagent collective behaviour and a few other cases. It does not work well for systems with prespecified goals, systems with funnels/bottleneck transitions, in short, for cases where you have to relinquish potential futures to commit to a decision. This needs to be treated with a different principle.

The information-theoretic treatment takes into account action routes with controllable dynamics, it will avoid state space regions with more noise and thus less controllability. Importantly: it is the "potential" to do something, it does not enforce to actually carry out the option. Having the option is all that counts. Like in chess, the "threat" is more effective than the "carrying out".

Disclaimer: I am one of the authors of the empowerment work.

Reply on news.ycombinator.com