Is APT still working to improve the quality of the Bots?
First, let me say that I am very pleased with the site, so please don't take this as a criticism.
My question: can we expect the AI of the bots to get better as more and more players join and play more and more hands? The reason that I ask is that even the toughest of the bots don't seem to be quite up to the level of the 1-3 games at, for example, the Wynn in LV.
The reason that I ask is that while this is a pretty good practice tool (I live in an area where neither on-line nor local card rooms are an option), I am quite sure that there will be a period of major adjusting to the extra player toughness once I get to LV (major adjusting = losing money). Ideally, I would like the Bots to be at least as good, and maybe even a little bit better than at least the 1-3 game at the Wynn/Aria/Orleans (with the ability to dial-down dial-up their toughness by choosing which room to play in (e.g. Phil's Garage (dial-down) or KGBs room (dial up).
Side note - I know that over the past few years chess AI has really gotten to the point where I can't beat it with any consistency. I think that Full ring poker is a lot tougher to "solve" than chess, but I would love it to get to that point someday soon. So that is really part of the reason that I am a subscriber to the site. Still I would like it to get to the point where if I can beat the bots at KGB (or even 1 level below), I can beat most of the 1-3 games in LV.
Hoping to hit the ground running.
Welcome
It looks like you're new here. If you want to get involved, click one of these buttons!
Comments
Thanks for the comments. I agree with you on all counts. AI is getting much better in many types of games, and poker is no exception. Our bots are pretty decent, but definitely not world beaters, which is why we are working on a pretty major reworking of them. I don't know the specific timing because major reworks take a long time, but I can tell you that we have been working on major changes to the bots for at least a year.
Just to give a little insight into why this is challenging - we want our bots to both simulate humans, but also be very good at the top levels. This is challenging. You may have noticed that we have some new tools in our tools and games tab that use ICM. So we have the capabilities now to make our bots play some pretty awesome heads up poker. However, most humans are not perfect and they suffer from some pretty significant biases and systematic leaks. So getting the bots to simulate this and also be good is a very tough task. Further, the best way to play is different in cash ring games, six max, sit n go, and of course MTT. So getting the bots good at ALL of these games is a monumental task.
But yes, I think you are correct that our bots are currently set to simulate decent but not great players. We are working to make them simulate players who could consistently win in a game like what you describe.
Where is the weak spot -- why would they lose, in particular?
Let's say a bot/android sat down at the above described live table, played one orbit (a full round starting at the big blind and ending before the blind came around again), and left. There would be no predictability/opportunity for opponents to pick up patterns and adjust. Would the bot be a lifetime loser doing this?
What's ICM?
If simulating human biases is difficult, can one of the AI characters be "I-Deal-Bot" (tm), the GTO-playing "ideal" opponent? I know GTO is a loaded term in full ring and/or no limit, but it'd still be cool to face an AI which was unbiased and optimized (if there is such a thing).
I also had the thought that a "sandbox" mode would be helpful, too. Sandbox mode would involve seeing the opponents' cards all the time and reloading your/their chips at will, and maybe setting MTT/Sit-n-go/Cash and #/selection of opponents as well. Even the ability to restart a hand with the same parameters (incl. AI biases) and holdings would be interesting. Of course, keeping records on this would be useless.
But the reason I bring up sandbox mode is that part of my learning process is figuring out how to obtain maximum extraction amounts from my good hands. Of course, working through it in this way could be distracting and/or counterproductive to the development of good instincts. Still, it's a thought.
I am curious if you guys have set up the bots to play each other. Would any of them beat the others with a significantly higher frequency?
"Just to give a little insight into why this is challenging - we want our bots to both simulate humans, but also be very good at the top levels. This is challenging. . . . However, most humans are not perfect and they suffer from some pretty significant biases and systematic leaks. So getting the bots to simulate this and also be good is a very tough task. . . . "
First, I want to thank you and your team for having an already outstanding product and I am excited to hear that there is a major reworking of the bots in progress!! I like having things to look forward to.
I am not sure that the bots need to simulate "real" players as much as to accurately simulate different player styles. I think that you can then create different levels of competency by increasing or decreasing the quantity and type of significant biases and systematic leaks. That would allow us players to move up the ladder of game toughness . . . finally arriving at poker perfection/nirvana - KGB's dungeon.
There are several different winning styles (e.g. loose-aggressive, tight-aggressive, maybe even some passive or uber aggressive styles as a counter/adjustment to certain aggressive or passive tendencies in the particular opponent). So having bots that play exceptionally well within those styles and then perhaps having a few than can randomly switch between those styles and/or switch as an adjustment to the game dynamics would probably make for a pretty tough game. Of course, I guess we don't want the bots to be so good that we can never be winners and then become overly discouraged. However, I do think that we would like it to always be a challenge, always require us to constantly be on our toes, and to punish us for our mistakes until we work them out. Perhaps the weekly training program could tell us at what level of competency we are currently at.
Needless to point out it's easy for me to say and damn difficult to actually design and build it. My guess is that we are several iterations away, but my guess is that you are the guys who can do it.
We have done that. That was the way we simulated the WSOP final table last year with Qui. Some of the bots are better than other ones. I think the aggressive ones tend to do better against the other bots,
Small technical bug:
In the Sit 'n Go simulation, I unchecked "Zip to End." If I fold but there is an all-in between the bots, the all-in percentages shown include mine along with theirs, even though I am out of the hand. This is obviously pretty inconsequential, but I figured it'd be worth giving you guys a heads-up about it.
It's interesting that the aggressive ones have the edge in bot vs. bot play. Do you think that that reflects specific aspects of the dynamics of bot play (especially bot vs. bot), or do you think that it's a natural mathematical property that plays we think of as "aggressive" or even over-aggressive are +EV over the long haul in games that are anywhere along the vast range of "typical?"
And are you referring to tournament play, cash play, or both?
The simulations must have been interesting. I would hazard a guess that some of the more aggressive ones ended up with a bimodal distribution in regards to placing, i.e.: maybe they either break away (upside) from the pack or bust out early. I would be curious to find out if this is or isn't true.
I think any strategy can work. The aggressive play winning is probably related to the bots, but I do think aggressive play is harder to adjust to in a real game. I would expect there is higher variance with aggressive play, but I haven't ever looked to see.
We have run bot vs bot in tournament play mostly.
I don't mean to necro an older thread, but I didn't see an update to this in any other forum posting. Have you rolled out the updated AI Bots yet?
In trying to determine what level membership commitment I want to make, it's important for me to get an idea of the continual added material. I see some of the new features you've rolled out lately, and that's very encouraging! So just following up on this one as well.
Hi,
Thanks for being a member - we appreciate it. And thanks also for posting on the forum because I like hearing from our members.
Short answer is that we will continue to add training materials and options to the site - Steve (my brother and co-founder of APT) and I love doing this and we've built up a great team here and a fantastic community of members.
Longer answer:
So our philosophy has changed since that thread almost a year ago. We ended up focusing on the Combat trainer and partnering with pros for webinars because the majority of the feedback we get on the bots are that with the right choice of settings, they can be made to play like almost any style of real human opponent. I definitely find they play a lot like the folks I play against in live games both in Florida and Vegas. I've become just an infinitely better poker player since we put this site together - which was the whole reason we started doing this computer-based training in the first place. Steve is and has been a great poker player and coach for a long time. I wanted a training site that would make ME a better player, and ten years ago when we started this nothing existed that came close to what I was looking for, so we built it ourselves.
I think it is important to recognize that we are not a content-driven site and we don't want to be that. Our entire training method is based off of large amounts of repetition using the game, while studying the odds and equity tools in the brain button, then working on the areas in the training plans. We also think the Combat trainer is going to prove to be an amazing way for players to plug leaks in some very specific settings that just can't be practiced in any way other than using our game. We're going to continue adding more and more of those scenarios. We're also going to continue to add Beat the Pro Challenges, which I think are akin to sweating a pro, which is definitely a great way to learn some new tricks. And of course we will keep doing webinars, adding new options to the game, etc.
Bottom line on the bots is that we made a decision that we don't want our bots to be GTO bots like Snowie, which is the direction we were heading before. That works great if you are playing extremely high buy-in poker and your opponents are world-beaters. We are focusing on a very different type of training - playing against the types of players you typically see at low to mid-level games, and the types you see interspersed with top players at high-level games. I've seen benefits to my game from using our site against players of all levels, and I think I wouldn't have improved as much if our bots played substantially better poker - it wouldn't match the typical humans we run into at the tables. We could spend a lot of time and make the bots more "perfect players", but I think that would end up making our game less realistic.
Anyway, thanks for bringing this question back because I'm happy to address our philosophy and hear other thoughts and comments. The directions we head with the site are almost completely driven by what our members tell us has been successful for them because in the end if the site is better for our members it is better for my game also. I feel like the guy from that hair club who in his ads always says that he's not just the founder, he's also a client. I very much match the profile of our "typical" member - a recreational player who is a professional at something other than poker, but wants to max out the poker experience when I do play and has dreams of someday winning a gold bracelet as an amateur lol.
Allen
Thank you very much for your informative reply and sharing your evolving philosophy! I've been away from poker for some time, and have now only recently returned and for the first time started playing in live cash games as opposed to strictly MTT. So I truly do think this program is exactly what I'm looking for. I'm certainly not looking for perfect player bots. By no means am I playing against them at the 1-2 NL Level. I look forward to using all the tools you have provided at APT!