DeepMind's AlphaZero plays chess like a tornado in the junkyard

[In situ]

That's exactly the "survival" I was talking about. I certainly did not mean to imply that the lengthy survival of any individual of a species was a "goal" of evolution.

I see, you understand it correctly but just express your understanding poorly. Well, in forum like this I tend to nitpick on everything because others, readers, might pick up the wrong understanding. And then I have to correct them as well.

 

[Tom 1]

Not sure what you mean with 'code', but my initial reaction is "no that is not how really how works", which tells me I problay did not understand you properly.

On a generic abstract level: the gradient in the fitness landscape.

If you are on a gradient you cannot avoid to move along the gradient. I don't like the picture of the fitness landscape with hills you need to climb. A hill suggest some kind of force are needed to push you up the hill, while rater the opoist is the case; you are falling down into a valley. The gradient works more like an energy minimum; if you put something in the empty air it will fall back to the ground because that is the direction in which the gravitational gradient points.

I don’t have the science or programming background to frame the question properly, it’s just something I’m curious about, and your answer makes sense. Like most threads this one seems to have gone off in a few different directions so I’m still not sure how this relates to the tornado analogy. Anyway.

 

[In situ]

Is it?

Yes. You just quoted it yourself:

As described in the link:
The focus of Monte Carlo tree search is on the analysis of the most promising moves, expanding the search tree based on random sampling of the search space.

This is designed and purposeful, a code that is designed for a particular type of purpose, and has a kind of intelligence in the form of analysis algorithms. It isn’t self generated, and when it has completed its purpose, it either does something else it was coded to do, or it stops. I assume you don’t think that random mutation is designed, so what would be the explanation for why it happens - not the process of how it happens, or the fact that it does happen, but why it does, what is the driving force behind it, in your view?

I think you miss the point; MCTS are designed to make their own designs.

I grant you the fact that a MCTS algorithm are designed by humans, but what a machines does which is equipped with an MCTS algorithm are not designed by human but the machines itself.

Analogous, as far as we can tell evolution is such design machine. I.e. the design process of evolution is a consequences of natural laws not an intelligent designer. An intelligent designer could possible be behind the laws of nature, but that is beside the point. It would be like observing a tornado assemble a Boeing 747 in a junkyard. Obviously the tornado did it, but the question who, if any, made a tornado capable of this miracle is another question.

 

[Tom 1]

Yes. You just quoted it yourself:






I think you miss the point; MCTS are designed to make their own designs.

I grant you the fact that a MCTS algorithm are designed by humans, but what a machines does which is equipped with an MCTS algorithm are not designed by human but the machines itself.

Analogous, as far as we can tell evolution is such design machine. I.e. the design process of evolution is a consequences of natural laws not an intelligent designer. An intelligent designer could possible be behind the laws of nature, but that is beside the point. It would be like observing a tornado assemble a Boeing 747 in a junkyard. Obviously the tornado did it, but the question who, if any, made a tornado capable of this miracle is another question.

Righht. I’m assuming you don’t mind being peppered with questions, as you keep responding. What is the understanding of natural laws as regards their application - I mean this seems analogous to some extent also, in that living organisms don’t choose to obey these laws as such, in the same way that we consciously do (or don’t). Is it that these laws are boundaries, or frameworks, or driving forces that herd organisms along a particular path, or an interaction of different forces that lead to particular results through a process that the laws allow, and any aberrations cease to be if they don’t fall within the limits of these laws, or something like that?

 

[In situ]

I don’t have the science or programming background to frame the question properly, it’s just something I’m curious about, and your answer makes sense.

That is fine, I will take that into account.

Like most threads this one seems to have gone off in a few different directions so I’m still not sure how this relates to the tornado analogy. Anyway.

The tornado is an analogy for how creationists perceive evolution; not possible. And everyone agrees, if evolution did in fact work that way it would not be possible. Much further than so you should not push the tornado analogy.

My point with the tornado analogy is this; I can point to machines which operates under similar condition as evolution, and they are capable of design, and in my view, without any designer being around telling them how to design things. So my question is; according to creationists, what is the fundamental differences between these machines and evolution? What make one (AlphaZero) work, but the other (evolution) not?

And as I pointed out in my OP: it does not matter if the machines has been designed to do this, what matter is why one machine work while another machine, operating under the very same principles, does not work.

 

[In situ]

Righht. I’m assuming you don’t mind being peppered with questions, as you keep responding.

So far it is okay.

What is the understanding of natural laws as regards their application - I mean this seems analogous to some extent also, in that living organisms don’t choose to obey these laws as such, in the same way that we consciously do (or don’t).

I would not consider anyone being able to breaking any laws of nature, not even we. What you refer to are agents which are able to run multiple simulation of alternative futures, based on the laws of nature, and pick the outcome which benefits its intention the most (btw this is what AlphaZero is doing in the Monte Carlo Tree Search). In these simulations an agent also need to take a probabilistic account of other active agents possible actions. Making things even more complicated.

This video partly exemplifies what I talk about:

TED talk on the same issue with some very nice demonstrations:

Obviously these simulations require some kind of deterministic world to even work otherwise the simulated outcomes, i.e. the predictions, cannot be assigned realistic probabilities and therefore are worthless to judge between. I.e. free will require reality to be predictable to some degree, i.e. the world needs to be coherent and understandable.

Is it that these laws are boundaries, or frameworks, or driving forces that herd organisms along a particular path, or an interaction of different forces that lead to particular results through a process that the laws allow, and any aberrations cease to be if they don’t fall within the limits of these laws, or something like that?

It seem to me that you conflate different concepts right now. I can try to nest them out, but right now I am a little bit tired and cannot think properly so I have probably not even understood your question correctly. With that disclaimer in mind, it looks like you on one hand are talking about evolution as a design process, the answer is then the gradient in the fitness landscape, and on another hand you area talking about the purpose, or intention of evolution, the answer is then the location of the hills in the fitness landscape.

Now these concepts are related since you cannot have a hill without a gradient. However, the reason for the existence of hills (e.g. predators exists, you need wings to fly) is not in general the same as the reason for why someone is moving up a hill (e.g. escaping a predator or catching a pray).

Also notice the fitness landscape is multi dimensional, in where each dimension represent a factor. For practical purposes research tries to limit the dimension to a handful, important, factors. You may call these factors "driving forces" but in reality we probably, I would guess, talk about thousands of factor, if not millions which all contribute with different weights at different times. The driving force is, to my best understanding, the combination of these factors.

But some times a driving force can be a random events which open up new niches for other to exploit, e.g. the K-T extinction or simple geographical displacements or just random walk or most of the time it seems that not much evolution take place at all. Biologist argue quite a lot about what drives evolution and I am not the one to tell. However the basic mechanism for evolution always remains the same: a change of allele frequency in a population.

 

[TerryWoodenpic]

Even a PC can play chess from the rules.
A super computer can do it faster and deeper. It is not anything like a tornado.
It is a logic machine.
However that does not make creationists an more logical themselves.

 

[TerryWoodenpic]

Even a PC can play chess from the rules.
A super computer can do it faster and deeper. It is not anything like a tornado.
It is a logic machine.
However that does not make creationists an more logical themselves.

Humans play chess very differently to machines, men do not calculate every possible move, they make intuitive shortcuts derived from experience. Machines do it by grinding data even down obviously blind alleys.

 

[Strathos]

Ummmmmm....SOMEbody had to "create" the robot and its "intelligence" from the gitgo.

Didn't they?

This argument seems irrelevant from my perspective. Against theistic evolution, it's not even an objection.

Against atheistic evolution, it merely switches the topic. I think every atheist would agree that in order for biological evolution to occur, the universe and the laws of physics must first exist. So if you are asking them where those things come from, that's a very good question, but irrelevant to the ToE, which only explains how life diversified after it already existed.

 

[Tom 1]

It seem to me that you conflate different concepts right now.

Yes, I didn't pursue any science subjects beyond secondary school, and that was a long time ago, so I just have a few vague notions.
I'll try and be clearer in what I'm asking - I don't think the bible contains anything that is intentionally scientific, nothing that attempts to offer an explanation regarding how things work in the physical universe, at least not in any detail. As I understand it though the general Christian world view is that God at the point of creation put into motion certain 'laws' that enable the universe to exist, function, and produce life, on our planet at least. I take the term 'natural law' to be analogous, i.e. in human society we have legal codes, laws, that we recognise and follow/obey but obviously particles, atoms, cells, etc etc are not aware of laws, they don't choose to follow them, so what is meant exactly by the analogy? In the scientific sense do natural 'laws' mean inviolable determinants of behaviour? Is the understanding that these determinants arise from the behaviour of matter in the universe or do they direct the behaviour of matter? Were these determinants of behaviour in some way built into the appearance of matter following the big bang and then guided the behaviour of that matter, or did the behaviour of matter follow a pattern that led to these determinants somehow becoming fixed? Or is it something else?

 

[Chesterton]

The Monte Carlo Tree Search used by AlphaZero wont even work if it isn't randomized so I do not understand what you mean when you say there is no randomness.

My understanding is that "random" is a misnomer. It's a term of art in computer programming. Yes, it certainly can appear unpredictable, and look random to humans, but so can falling snowflakes. I would need to be convinced that electricity can do anything other than follow the laws of physics.

Stockfish has been given the rule of chess and heuristic tree searches and strategics discovered by humans etc etc in order to be able to play chess, i.e. Stockfish has been designed by humans, over several decades, to play good chess.

On the other hand, AlphaZero has only been given the rules and then discovered, by itself, how to play chess, i.e. AlphaZero was not designed to play chess, yet AlphaZero can play chess and it achieved superhuman capability with randomness and no plan or goal whatsoever in just a few hours, i.e. just like evolution.

Btw, AlphaZero is a game changer. Before AlphaZero, humans had to teach chess engines how to play chess. However, AlphaZero was has not been taught by humans to play chess, instead it seems like AlphaZero might actually teach human how to play chess instead. I.e. AlphaZero is better at chess than any human - ever.

If it's not intelligent design, then you tell me what accounts for the difference between the two programs. As someone noted earlier, if you're trying to make some point for atheistic evolution, you're undercutting your own position.

 

[Petros2015]

if randomness cannot create design, no matter what time is given, then what did cause AlphaZero to achieve a superhuman performance in chess in less than four hours time?

AlphaZero was told that there was a game by it's Creators, and it was told what constituted 'winning'. It was given feedback to know that it had won or lost. If it hadn't been given that feedback, it wouldn't have known how to improve. Randomness would remain just that, random, instead of steps to perfection. Even if it randomly created something good, it would just throw it away, unable to evaluate it. The capacity for evaluation is essential. In order to improve, you have to have a standard to improve against.

Likewise for humans, we have a capacity for evaluation and a standard to improve against. This falls at a variety of levels, from basic survival to moral choices and spirituality. If there was just life on the physical level, I would think that life would evolve to the point where it was effectively immortal and call it a day. We really haven't seen that yet. Why is there such diversity? Soup should beget better soup, and eventually immortal soup. Not birds.

 

[FrumiousBandersnatch]

A will kind of implies consciousness. Perhaps with 'will' you are referring to 'intentions'? There are machines, i.e. agents, which have intentions. Intention does not require and agent to be consciousness as far as I understand, but you may judge it differently than me and in some cases I've started to doubt myself weather or not we already made conscious robots....

You be surprised what AI can do these day then...

However I think the claims made are hyped. I read their research paper and I think the conclusion the robots passed the self-awareness test is flawed - for many reasons. Anyway the point is that these robots demonstrates intention.

I don't think any robots are recognisably conscious as we generally think of consciousness, but there are some that have specific aspects or components of consciousness, however these are typically fairly task-specific, not well-integrated, and only weakly generalizable.

Some robots, like Vestri and the iCub series are designed to learn for themselves and by example, in a way conceptually similar to biological brains. This kind of system is most likely to produce patterns of behaviour that are familiar to us, but unlikely to produce recognisable consciousness without the provision of facilities for metacognition, narrative generation, self-modeling, and so-on.

Part of the problem is that consciousness itself is poorly-defined - it appears to consist of a bunch of interacting capabilities, few of which seem essential, but which together give the sense of an integrated coherent self.

And if you haven't seen what Boston Dynamics are up to, this might impress on you.

These robots are not just acting autonomous, they are also re-acting to the unpredictable word around them.

Boston Dynamics seem to have captured the algorithms and heuristics that animals naturally use to control their movement and balance - I've not seen any other robots come close to their level of reflexive adaptability - the gross bodily movements are impressive, but the small things, like the way they can casually push open a door and hold it open with one arm as they walk through, seem uncannily human - makes me wonder whether they used motion capture templates from humans to generate their control sequences...

 

[FrumiousBandersnatch]

My understanding is that "random" is a misnomer. It's a term of art in computer programming. Yes, it certainly can appear unpredictable, and look random to humans, but so can falling snowflakes. I would need to be convinced that electricity can do anything other than follow the laws of physics.

Strictly speaking, this is true (except that quantum interactions do appear to be truly stochastic). But in general use, randomness is effectively unpredictability.

So in program code, repeatable pseudorandom numbers are generated algorithmically from arbitrary numerical seeds (e.g. the time between user inputs), and, where a higher level of randomness is required, random numbers can be generated by amplifying thermal noise in an electronic circuit - the same principle has been found in the brain, where neural circuit noise is used to boost perception of threshold auditory signals, and to provide arbitrary timings when unpredictability is required.

For ultimate randomness, measuring radioactive decay is thought to give truly stochastic results, as it's a quantum process.

 

[FrumiousBandersnatch]

AlphaZero was told that there was a game by it's Creators, and it was told what constituted 'winning'. It was given feedback to know that it had won or lost. If it hadn't been given that feedback, it wouldn't have known how to improve.

As I understand it, unlike previous Alpha versions, AlphaZero had no external feedback or reinforcement. It was given the rules of the game and what constituted a win, loss, or draw. From that it could evaluate its own performance.

If there was just life on the physical level, I would think that life would evolve to the point where it was effectively immortal and call it a day. We really haven't seen that yet. Why is there such diversity? Soup should beget better soup, and eventually immortal soup. Not birds.

Evolution by natural selection produces a diversity of creatures, particularly when there is a diversity of environments, as the genetic makeup of populations changes over time; and there are a few creatures that are effectively immortal (if not killed by external forces).

 

[durangodawood]

....despite we got a "Boeing 747" in front of our eyes created by a "tornado"?

I dont think thats what happened.

The learning algorithm is in no way comparable to a tornado. AlphaZero has a built in 'telos' and a built in method to achieve it. Tornados lack that.

 

[Speedwell]

I kind of lost you there, I am not used to think in terms of such elaborate colorful language. Don't take me wrong, I think it was beutiful written, and I can sense it carries a lot of meaning , but it was written far beyond my own comprehension.

1. Evolution proceeds by random variation followed by selection, over and over.
2. Some Christians have concluded that the 'random' part of the process must, of necessity, break any possible causal chain between God and the product of selection.
3. This conclusion seems more likely to be reached by Christians in denominations which lack a well-developed formal theology--fundamentalist Protestant denominations for the most part.

 

[Speedwell]

I don’t have the science or programming background to frame the question properly, it’s just something I’m curious about, and your answer makes sense. Like most threads this one seems to have gone off in a few different directions so I’m still not sure how this relates to the tornado analogy. Anyway.

The "tornado" analogy is not In situ's; it was coined by the famous scientist and evolution critic Fred Hoyle many years ago.

Although it has become very popular with creationists--who push it up our noses constantly--when examined closely it fails as an analogy for describing evolution. It fails as an analogy for the machine learning In situ is describing as well, for similar reasons, and I suspect that In situ only brought it into the discussion out of a sense of irony.

 

[In situ]

AlphaZero was told that there was a game by it's Creators, and it was told what constituted 'winning'. It was given feedback to know that it had won or lost.

AlphaZero was told how the world works, i.e. the laws, but not told how to act in the world. It was the latter AlphaZero discovered by itself, i.e. AlphaZero discovered how to play chess. Previous chess program been told both because they cannot discover how to play chess..

It was given feedback to know that it had won or lost. If it hadn't been given that feedback, it wouldn't have known how to improve.

I agree, the research paper agree, you agree, everyone agree on this. But you have to do this. If there are no laws, i.e. restriction then anything goes and you have chaos. It is a kind of pointless world. What are you suposed to learn in this world? Not to play chess that is for sure. Therefore you need rules as a base to define the world in order for there to be something to discover.

Randomness would remain just that, random, instead of steps to perfection. Even if it randomly created something good, it would just throw it away, unable to evaluate it. The capacity for evaluation is essential. In order to improve, you have to have a standard to improve against.

Again, I agree and that is why you need rule to start with to define how the world works. I do not see how you disagree with me on anything, rather it seams to me you have conflated to different concepts into one; how the world works v.s. how to act in the world. These are not the same things. and there is no general rule for how to act in the world. AlphaZero discovered the same way to play chess as human has, and it seams it discovered some new ways as well in the process. In other words, instead of humans telling AlphaZero how to play chess, AlphaZero is telling humans how to play chess.

The point is AlphaZero has designed something, which no human has designed - ever!

Likewise for humans, we have a capacity for evaluation and a standard to improve against. This falls at a variety of levels, from basic survival to moral choices and spirituality. If there was just life on the physical level, I would think that life would evolve to the point where it was effectively immortal and call it a day. We really haven't seen that yet. Why is there such diversity? Soup should beget better soup, and eventually immortal soup. Not birds.

Your point is?

 

[Speedwell]

If there was just life on the physical level, I would think that life would evolve to the point where it was effectively immortal and call it a day.

It has. Individual creatures are not immortal; species are not immortal, but life goes on.