Artificial Intelligence and the stock market

Discuss all general (i.e. non-personal) investing questions and issues, investing news, and theory.
Workable Goblin
Posts: 68
Joined: Fri Mar 01, 2019 8:37 pm
Location: Honolulu, HI

Re: Artificial Intelligence and the stock market

Post by Workable Goblin » Sat Oct 12, 2019 11:32 am

tadamsmar wrote:
Sat Oct 12, 2019 6:15 am
Another way that separate neural networks can be "merged" is by having them vote or collaborate in some manner. Humans do stuff like that to combine their knowledge. Market pricing is a (possibly imperfect) mechanism that combines the knowledge and experience of many humans, and these days it probably also combines the knowledge and experience of AI entities. Also, there might be multiple (not fully integrated) neural networks collaborating in our own brains.
This is actually a much older non-machine learning technique that is mostly used in highly critical applications to provide redundancy against component failures (e.g., the Space Shuttle used this technique so that failures of any of its computers could be compensated for). The problem in this context is that "voting" doesn't actually allow any of the components to improve and learn from any "accidents and incidents," as you might put it, and it may very well be the case that the network that "knows" the correct answer is outvoted by the ones that don't. So, while I don't work at Waymo (or, indeed, any of the other self-driving car companies) I feel pretty confident in stating that they don't use this technique for "merging" the experience of their different cars. They might use it in the car hardware so that component failures don't cause the self-driving systems to shut down, of course (in fact, cars with a safety driver do use it)

jibantik
Posts: 207
Joined: Fri Nov 24, 2017 1:05 pm

Re: Artificial Intelligence and the stock market

Post by jibantik » Sat Oct 12, 2019 9:05 pm

finite_difference wrote:
Fri Oct 11, 2019 11:09 pm
Do either of you play Go? It is a sincere question, because I think you are vastly underestimating the complexity of the game. I thought it looked simple too until I started playing.

Has Google really put the same amount of effort it put into defeating Go grandmaster Ke Jie, toward defeating one of the top DOTA or LoL players, and come up empty handed? If not, I don’t think it’s a fair comparison.
I have played some, I'd say its simple to learn but takes a lifetime to master. The solution space is massive, but it's much more constrained than DOTA and easier to model.

I know with DOTA, openAI spent years researching and training AI to play the game and the result was incredibly unimpressive to me. And this was trained with supercomputers that played 180 years worth of games PER DAY for months.

Even I was starting to somewhat believe the hype of AI, but witnessing AI playing DOTA, it was clear that with our current technology we are still at a superficial level. The AI was not "smart" and could not adapt and looked downright stupid/random when the game didn't progress as it needed to. Not to mention, they played the AI on a restricted version of the game so it would be better.

General intelligence is a long ways out from our current tech.

randomguy
Posts: 8333
Joined: Wed Sep 17, 2014 9:00 am

Re: Artificial Intelligence and the stock market

Post by randomguy » Sat Oct 12, 2019 10:23 pm

jibantik wrote:
Sat Oct 12, 2019 9:05 pm
finite_difference wrote:
Fri Oct 11, 2019 11:09 pm
Do either of you play Go? It is a sincere question, because I think you are vastly underestimating the complexity of the game. I thought it looked simple too until I started playing.

Has Google really put the same amount of effort it put into defeating Go grandmaster Ke Jie, toward defeating one of the top DOTA or LoL players, and come up empty handed? If not, I don’t think it’s a fair comparison.
I have played some, I'd say its simple to learn but takes a lifetime to master. The solution space is massive, but it's much more constrained than DOTA and easier to model.

I know with DOTA, openAI spent years researching and training AI to play the game and the result was incredibly unimpressive to me. And this was trained with supercomputers that played 180 years worth of games PER DAY for months.

Even I was starting to somewhat believe the hype of AI, but witnessing AI playing DOTA, it was clear that with our current technology we are still at a superficial level. The AI was not "smart" and could not adapt and looked downright stupid/random when the game didn't progress as it needed to. Not to mention, they played the AI on a restricted version of the game so it would be better.

General intelligence is a long ways out from our current tech.
I am curious when you last saw the OpenAI Dota player? My understanding is the 2018 player was pretty weak. The 2019 one has been beating everyone at the subset of the game it plays. I will not hazard a guess at how much OpenAI has spend on it versus how much google spent both in terms of time and money.

My comment wasn't that go was a simple game. It was that go was a simple problem.What you need to know and the problems you have to do to play go well is vastly simpler than driving a car and so on.

User avatar
tadamsmar
Posts: 8521
Joined: Mon May 07, 2007 12:33 pm

Re: Artificial Intelligence and the stock market

Post by tadamsmar » Sun Oct 13, 2019 10:03 am

Maybe I am missing something, but the discussion DOTA seems to be almost 6 months out of date. AI crushed the international champ team in April:

https://www.theverge.com/2019/4/13/1830 ... l-champion

And here is a recent article where a poker bot beats multiple pros:

https://www.wired.com/story/new-poker-b ... iple-pros/

And poker sites have been trying to ban bots for quite a while:

https://www.cardschat.com/news/partypok ... ites-83960

Humans talk about how 7 human years = 1 dog year. AI may soon be talking about how 7 AI learning days = 1 human year. Or 7 minutes, or 7 seconds, or...

With computer capabilities, what was true a few months ago is often false now.

Topic Author
masonstone
Posts: 472
Joined: Thu Feb 01, 2018 4:01 pm

Re: Artificial Intelligence and the stock market

Post by masonstone » Wed Oct 16, 2019 12:49 am

tadamsmar wrote:
Sun Oct 13, 2019 10:03 am
Maybe I am missing something, but the discussion DOTA seems to be almost 6 months out of date. AI crushed the international champ team in April:

https://www.theverge.com/2019/4/13/1830 ... l-champion

And here is a recent article where a poker bot beats multiple pros:

https://www.wired.com/story/new-poker-b ... iple-pros/

And poker sites have been trying to ban bots for quite a while:

https://www.cardschat.com/news/partypok ... ites-83960

Humans talk about how 7 human years = 1 dog year. AI may soon be talking about how 7 AI learning days = 1 human year. Or 7 minutes, or 7 seconds, or...

With computer capabilities, what was true a few months ago is often false now.
Hence the AI explosion. AI will grow exponentially once it learns how to program and engineer itself. This thus will lead to monopolies (similar to how google is the search engine monopoly now).

shess
Posts: 237
Joined: Wed May 17, 2017 12:02 am

Re: Artificial Intelligence and the stock market

Post by shess » Wed Oct 16, 2019 1:00 am

masonstone wrote:
Wed Oct 16, 2019 12:49 am
tadamsmar wrote:
Sun Oct 13, 2019 10:03 am
With computer capabilities, what was true a few months ago is often false now.
Hence the AI explosion. AI will grow exponentially once it learns how to program and engineer itself. This thus will lead to monopolies (similar to how google is the search engine monopoly now).
Maybe, and maybe not. When technology lands, everyone predicts crazy things like flying nuclear cars and space lasers, but what we actually end up with are light-up shoes and Blue-ray disks. As of yet, nobody has discovered any sort of generalized AI, it's all still in very specialized domains. Waymo and other AI-driver companies are still approaching it by simply overwhelming the problem with data, but AI drivers still don't comprehend why they shouldn't run over children or how friction works. It's entirely possible that by the time we build an AI which is as general-purpose as a human, it will also be as unreliable as a human and hardly any faster than a human.

Don't get me wrong, I've been agog about the Singularity for a few decades, and I do think we'll probably eventually manage a generalized AI. I'm just not going to change any of my current plans because of it. Remember the Fifth Generation Project and Prolog? Yeah...

User avatar
tadamsmar
Posts: 8521
Joined: Mon May 07, 2007 12:33 pm

Re: Artificial Intelligence and the stock market

Post by tadamsmar » Wed Oct 16, 2019 10:57 am

masonstone wrote:
Wed Oct 16, 2019 12:49 am
Hence the AI explosion. AI will grow exponentially once it learns how to program and engineer itself. This thus will lead to monopolies (similar to how google is the search engine monopoly now).
Monopolies are not new. Standard Oil was probably the biggest US monopoly ever as a percentage of GNP and it was going stong 130 years ago.

Perhaps AI will be effectively applied to doing a better job of protecting monopolies from laws and regulations.

randomguy
Posts: 8333
Joined: Wed Sep 17, 2014 9:00 am

Re: Artificial Intelligence and the stock market

Post by randomguy » Wed Oct 16, 2019 11:55 am

tadamsmar wrote:
Sun Oct 13, 2019 10:03 am
Maybe I am missing something, but the discussion DOTA seems to be almost 6 months out of date. AI crushed the international champ team in April:

https://www.theverge.com/2019/4/13/1830 ... l-champion

And here is a recent article where a poker bot beats multiple pros:

https://www.wired.com/story/new-poker-b ... iple-pros/

And poker sites have been trying to ban bots for quite a while:

https://www.cardschat.com/news/partypok ... ites-83960

Humans talk about how 7 human years = 1 dog year. AI may soon be talking about how 7 AI learning days = 1 human year. Or 7 minutes, or 7 seconds, or...

With computer capabilities, what was true a few months ago is often false now.
On the other hand, look at what Waymo's self driving cars were doing back in 2012. It is almost 8 years later and while they have made progress, they are learning slower than a human. An 10 year old driving like that in 2012 would have a license by now:)

Yes the bots are getting better. In DOTA this is a debate about how much is because of the limited game (i.e. humans aren't used to playing it), how much is just better micro (AI use APIs not machine vision and mouse/keyboard controls to play the game. It is a bit of nitpicking), and how much is just better strategy but the improvement in the last year has been really solid. In poker, bots have been good enough to beat the average player for a long time. Sites aren't worried about a bot sitting down at some 100/200 table and beating the pros. They are worried about someone running 1000 bots and sitting down at the .25/.50 tables are murdering the horrible players.

General AI is sort of a red herring. Yeah it would be a game changer. But specialized AI's alone could be killer in a lot of fields in term of replacing skilled and semiskilled labor. But guessing what progress is made and when is almost impossible.

finite_difference
Posts: 1455
Joined: Thu Jul 09, 2015 7:00 pm

Re: Artificial Intelligence and the stock market

Post by finite_difference » Thu Oct 17, 2019 10:37 am

jibantik wrote:
Sat Oct 12, 2019 9:05 pm
finite_difference wrote:
Fri Oct 11, 2019 11:09 pm
Do either of you play Go? It is a sincere question, because I think you are vastly underestimating the complexity of the game. I thought it looked simple too until I started playing.

Has Google really put the same amount of effort it put into defeating Go grandmaster Ke Jie, toward defeating one of the top DOTA or LoL players, and come up empty handed? If not, I don’t think it’s a fair comparison.
I have played some, I'd say its simple to learn but takes a lifetime to master. The solution space is massive, but it's much more constrained than DOTA and easier to model.

I know with DOTA, openAI spent years researching and training AI to play the game and the result was incredibly unimpressive to me. And this was trained with supercomputers that played 180 years worth of games PER DAY for months.

Even I was starting to somewhat believe the hype of AI, but witnessing AI playing DOTA, it was clear that with our current technology we are still at a superficial level. The AI was not "smart" and could not adapt and looked downright stupid/random when the game didn't progress as it needed to. Not to mention, they played the AI on a restricted version of the game so it would be better.

General intelligence is a long ways out from our current tech.
Based on the following posts indicating that tremendous progress had been made over a year ago, does this change your mind?
The most precious gift we can offer anyone is our attention. - Thich Nhat Hanh

jibantik
Posts: 207
Joined: Fri Nov 24, 2017 1:05 pm

Re: Artificial Intelligence and the stock market

Post by jibantik » Thu Oct 17, 2019 10:05 pm

finite_difference wrote:
Thu Oct 17, 2019 10:37 am
jibantik wrote:
Sat Oct 12, 2019 9:05 pm
finite_difference wrote:
Fri Oct 11, 2019 11:09 pm
Do either of you play Go? It is a sincere question, because I think you are vastly underestimating the complexity of the game. I thought it looked simple too until I started playing.

Has Google really put the same amount of effort it put into defeating Go grandmaster Ke Jie, toward defeating one of the top DOTA or LoL players, and come up empty handed? If not, I don’t think it’s a fair comparison.
I have played some, I'd say its simple to learn but takes a lifetime to master. The solution space is massive, but it's much more constrained than DOTA and easier to model.

I know with DOTA, openAI spent years researching and training AI to play the game and the result was incredibly unimpressive to me. And this was trained with supercomputers that played 180 years worth of games PER DAY for months.

Even I was starting to somewhat believe the hype of AI, but witnessing AI playing DOTA, it was clear that with our current technology we are still at a superficial level. The AI was not "smart" and could not adapt and looked downright stupid/random when the game didn't progress as it needed to. Not to mention, they played the AI on a restricted version of the game so it would be better.

General intelligence is a long ways out from our current tech.
Based on the following posts indicating that tremendous progress had been made over a year ago, does this change your mind?
It does not :D .

Previous poster summed it up well, but I will reiterate a couple things about the DOTA one since I am familiar with that.
  • The AI was playing on a vastly constrained version of the game, not the full game. It would have been terrible at the full game. Also, this hurt the pros who weren't accustomed to the constrained game.
  • The AI arguably had faster than human reaction times, better than human micro management ability, perfect accuracy on when timers, etc. These things made the AI stronger in team fights simply because it had faster reflexes, not because it had a "smarter" strategy.
  • I read that one of the pros was throwing the game in the April match and not really trying that hard (I didn't watch this game so may or may not be true).
  • It only played 2 games with the pros in the April match. Even the previous iteration could beat the pros _initially_, but eventually the pros saw what it was doing and adjusted and destroyed the AI because the AI could not adjust. I'd be willing to bet if the AI played the pro team more, it would win the first 5-10 games and then the pros would win every game after that.
The last point is why I found the AI particularly underwhelming.

randomguy
Posts: 8333
Joined: Wed Sep 17, 2014 9:00 am

Re: Artificial Intelligence and the stock market

Post by randomguy » Thu Oct 17, 2019 10:45 pm

jibantik wrote:
Thu Oct 17, 2019 10:05 pm

It does not :D .

Previous poster summed it up well, but I will reiterate a couple things about the DOTA one since I am familiar with that.
  • The AI was playing on a vastly constrained version of the game, not the full game. It would have been terrible at the full game. Also, this hurt the pros who weren't accustomed to the constrained game.
  • The AI arguably had faster than human reaction times, better than human micro management ability, perfect accuracy on when timers, etc. These things made the AI stronger in team fights simply because it had faster reflexes, not because it had a "smarter" strategy.
  • I read that one of the pros was throwing the game in the April match and not really trying that hard (I didn't watch this game so may or may not be true).
  • It only played 2 games with the pros in the April match. Even the previous iteration could beat the pros _initially_, but eventually the pros saw what it was doing and adjusted and destroyed the AI because the AI could not adjust. I'd be willing to bet if the AI played the pro team more, it would win the first 5-10 games and then the pros would win every game after that.
The last point is why I found the AI particularly underwhelming.
It some extent we are arguing about a minor point. It is pretty clear that given time and computer power, you can use this technique to make a DOTA bot that is unbeatable. Might take another 5 years of learning but that will be the end result. You can look at every time the best humans lose to a machine for the first time and you get the same excuses. Machine played oddly and surprised the humans. Humans didn't take the game seriously. And so on. It tends not to take long to progress to the next step and admitting the computers are just better.

It is clear that automation will keep progressing and doing things better than humans do. But I doubt there will be some singularity point where everything changes. Right now every problem requires training (i.e. your strawberry picking robot will not also pick blueberries) which slows down rollouts. But you should expect over time machines doing more and more work.

jibantik
Posts: 207
Joined: Fri Nov 24, 2017 1:05 pm

Re: Artificial Intelligence and the stock market

Post by jibantik » Thu Oct 17, 2019 11:54 pm

randomguy wrote:
Thu Oct 17, 2019 10:45 pm
It some extent we are arguing about a minor point. It is pretty clear that given time and computer power, you can use this technique to make a DOTA bot that is unbeatable. Might take another 5 years of learning but that will be the end result. You can look at every time the best humans lose to a machine for the first time and you get the same excuses. Machine played oddly and surprised the humans. Humans didn't take the game seriously. And so on. It tends not to take long to progress to the next step and admitting the computers are just better.

It is clear that automation will keep progressing and doing things better than humans do. But I doubt there will be some singularity point where everything changes. Right now every problem requires training (i.e. your strawberry picking robot will not also pick blueberries) which slows down rollouts. But you should expect over time machines doing more and more work.
I don't think the first point is clear at all. In fact, I'd say it's unlikely the case without some radically new techniques or technology advances. The second point I 100% agree with.

TomCat96
Posts: 790
Joined: Sun Oct 18, 2015 12:18 pm

Re: Artificial Intelligence and the stock market

Post by TomCat96 » Fri Oct 18, 2019 1:25 pm

It's important to distinguish among the disparate AI techniques out there.

As a hobby of mine, I like to tinker with OpenAI gym and manipulate some of the algorithms for these game bots.
This particular field of AI is called reinforcement learning as oppose to say your typical linear classifiers, LSTMs, autoencoders, and convolutional neural networks.

Of all these, linear classifiers are probably the best understood. And they do a decent job classifying things.
The mathematics of a linear classifier is essentially solving for a vector in an N-dimensional space, or minimization of an entropy function in steps so that the parameters of a neural network will gradually fit the data.

The mathematics is quite different for reinforcement learning, where the target is to search for an optimal policy.
There exists a concept in reinforcement learning that does not exist in your typical linear classifiers and convolutional neural networks--that of the Agent.

The Agent is an actor which may interact in a State space, and may produce an action. There exists a set of rewards (which is really a giant matrix).
The agent may act within this state space in discrete time steps. The objective is to find a policy, the set of actions, or subsequent states, or action, state tuples, over the course of the entire state space which maximizes the reward.

In my subjective opinion, reinforcement learning, the AI of bots, is an order of magnitude more complex than linear classifiers. Papers on reinforcement learning get posted regularly. But by and large, it's not the coding that is difficult. Code, and even libraries implementing the new mathematical techniques are available within days of the papers being posted. These libraries are available for use on Tensorflow.

Thus, tinkering with a new bot to interact with the OpenAI gym world can be just a matter of a few calls to some libraries. Truly Amazing really.

I have a bot of sonic running around interacting with his world, trying to run loops and collect gold rings. Believe it or not, its actually beaten the first stage once or twice.

Here's the problem though. How do I make the bot better?

I don't have a single clue. And I'm not sure anyone else does either. Sure you can tinker with a few things, combine AI techniques, tweak reward functions, throw more computing power at it. But we don't actually know what will make it better. If I throw 10 GPUs at it instead of 5, how much "smarter" will Sonic be? There are no metrics to determine this. Even for certain RL algorithm which mathematically guarantee convergence, the state space of these games are so ridiculously intractable to brute force that it's just not going to make much of a difference.

I won't speak for the other AI techs, but for DOTA botting it's not a matter of engineering at this point as it is going to the whiteboard. We can't engineer our way into better bot AI yet--at least not fundamentally. I would argue at this moment, it's really in the hands of the PhDs right now.

randomguy
Posts: 8333
Joined: Wed Sep 17, 2014 9:00 am

Re: Artificial Intelligence and the stock market

Post by randomguy » Fri Oct 18, 2019 7:15 pm

jibantik wrote:
Thu Oct 17, 2019 11:54 pm


I don't think the first point is clear at all. In fact, I'd say it's unlikely the case without some radically new techniques or technology advances. The second point I 100% agree with.
I suppose it is debateable but look at what they have done in 2 years. They go from playing a crippled 1 player game, to player a game where they lose agains the really good players, to beating the best players in the world. From reading the paper they aren't do much in the lines of new algorithms. More just tuning the algorithms to be more scalable so they can do more learning. Maybe this is as good as it gets but I expect when you have 10x the computer power and another 5 years of learning, the outcome will not be close. Past history has shown once the machines are competitive with humans, they quickly surpass them. The only way I see this not happening is if nobody spends the money.

It should be noted how limited these agents are. You spend 5 million dollars building a world class DOTA 2 bot. That bot can't play . DOTA or DOTA 3 (when/if it ships). The techniques will transfer but you will need to spend more millions on CPU time training for each occurence. That is pretty limiting. On the other hand a human that is got at Dota 2, will be decent at Dota 1 and would probably play well in Dota 3. On the other hand we are solving problems that we didn't know how to attack 20 years ago. If in 2000 you asked someone if a computer would be the best go player in the world in under 20 years, not many people would have taken the bet.

It is easy to both get carried away with optimism (everything will change) and pessimism (still can't do X) when looking at AI.

jibantik
Posts: 207
Joined: Fri Nov 24, 2017 1:05 pm

Re: Artificial Intelligence and the stock market

Post by jibantik » Fri Oct 18, 2019 10:44 pm

randomguy wrote:
Fri Oct 18, 2019 7:15 pm
jibantik wrote:
Thu Oct 17, 2019 11:54 pm


I don't think the first point is clear at all. In fact, I'd say it's unlikely the case without some radically new techniques or technology advances. The second point I 100% agree with.
I suppose it is debateable but look at what they have done in 2 years. They go from playing a crippled 1 player game, to player a game where they lose agains the really good players, to beating the best players in the world. From reading the paper they aren't do much in the lines of new algorithms. More just tuning the algorithms to be more scalable so they can do more learning. Maybe this is as good as it gets but I expect when you have 10x the computer power and another 5 years of learning, the outcome will not be close. Past history has shown once the machines are competitive with humans, they quickly surpass them. The only way I see this not happening is if nobody spends the money.

It should be noted how limited these agents are. You spend 5 million dollars building a world class DOTA 2 bot. That bot can't play . DOTA or DOTA 3 (when/if it ships). The techniques will transfer but you will need to spend more millions on CPU time training for each occurence. That is pretty limiting. On the other hand a human that is got at Dota 2, will be decent at Dota 1 and would probably play well in Dota 3. On the other hand we are solving problems that we didn't know how to attack 20 years ago. If in 2000 you asked someone if a computer would be the best go player in the world in under 20 years, not many people would have taken the bet.

It is easy to both get carried away with optimism (everything will change) and pessimism (still can't do X) when looking at AI.
Just one thing, a bot has never come close to beating pros at DOTA. It has beaten pros at an extremely constrained version of DOTA. Saying otherwise would be like a bot beating pros at Go on a 9x9 board and saying it beat pros at Go.

Also, i'd bet a lot of money that if you rolled back those games 100 times, even on this constrained version, the bot may win 5-10 games and the pros could win every game thereafter.

User avatar
tadamsmar
Posts: 8521
Joined: Mon May 07, 2007 12:33 pm

Re: Artificial Intelligence and the stock market

Post by tadamsmar » Wed Oct 23, 2019 10:27 am

A few weeks ago Morgan Stanley knocked 40% off the estimated value of Waymo, Google's self-driving car subsidiary, because they are still using safety drivers for their Phoenix pilot taxi service.

After that event, Google notified it's pilot riders that safety drivers would start to be eliminated soon. And Google complained about how the media has been hyping self-driving cars:

https://www.cnbc.com/2019/10/23/alphabe ... -cars.html

PS: I used the more familiar Google for the more correct "Alphabet".

Post Reply