Quote:
Originally Posted by Rapidesh123
You don't forget programming.
You may forget how to code, how to program in a language, but you never forget how to make algorithms.
if you knew how to program, you would understand that in order to make only the logic of a bot that plays 9/5 nit poker will take you more than 3k lines of code.
Now imagine making a bot that plays semi-GTO lol
Okay, since you mention it, let's consider GTO strategies first. Let's start with basics. John Nash proved that a GTO strategy
exists for any finite poker game with any number of players. But existence is the "easy" part. Finding a GTO strategy (or strategies) is the hard part.
Generally speaking there are two different but related tasks associated with GTO poker strategies. One is to "find" a GTO strategy, and the other is to "encode" a GTO strategy once found.
To date, the only game that has been "solved" is HULHE. And, truthfully, I believe that it is only essentially solved since there is no way of knowing whether a true equilibrium GTO strategy (set of strategies, one for SB and one for BB) has been found. But let's leave that to the side for our purposes here.
The good folks at the University of Alberta solved heads-up limit holdem a couple of years ago. They used a sophisticated "learning" algorithm and immense computing power to essentially have a computer play trillions and trillions of hands against itself to eventually arrive at an equilibrium strategy (pair).
Once found, the Alberta team encoded the GTO strategy into a computer called Cepheus that is available for the public to play against. The complete strategy takes tons and tons of storage space on a very large computer and consists of literally zillions of lines of code.
Now fast forward a year or two. The team at CMU developed new sophisticated "learning" techniques to have a computer approximate a HUNLHE GTO strategy. The program/computer is called Libratus. Libratus played trillions and trillions of hands of 200bb deep heads-up no-limit holdem against itself until the CMU team was confident that it had reached a fairly good GTO approximation (abstracting bet sizes).
Rather than encode the complete quasi-GTO strategy it developed, the CMU team stores the preflop and flop strategy elements in computer storage. For the turn and river, Libratus runs a separate solver to derive the optimal strategy given the action to that point. For the Libratus challenge, the CMU team rented time on a supercomputer for this purpose.
When asked if the Libratus program could be transferred over to a PC, the main developer said that it is theoretically possible but would take a lot of work. Once successfully converted to a PC-friendly form, on one of today's fast PC's, Libratus would probably take at least several minutes on its turn/river decisions as compared to 15-20 seconds on the supercomputer.
The point is that nobody initially sits down to "program" a GTO strategy. That concept, as stated, is meaningless. First you must find a GTO strategy (quasi). This itself is currently beyond the capabilities of all but a handful of teams of academic researchers. Once found it would then take either immense computer storage to store the strategy or immense computing power to calculate aspects of it on the fly (or both).
Should the above fill online poker players with glee? Of course not. Computing power is going to continue growing in leaps and bounds ala Moore's Law. So even if HULHE is the only game truly solved to date and even if a quasi-GTO strategy has recently been developed for only 200bb deep HUNLHE, these methods and techniques may well be used to develop quasi-quasi GTO solutions of other poker variants in some unknown time horizon.
But that is not the most important or salient fear that online poker players have regarding bots. Pre- and post-flop solvers already exist and their approximate solutions can be ported to a poker-playing bot. Far-from-perfect strategies can be developed fairly easily and surely can beat many low-to-mid stakes games.
And the same advances in computing power and storage which will improve the hunt for GTO strategies will mean that "far-from-perfect" computer-playing bots will also continue to improve in strength, speed, complexity, etc. This is surely the largest concern.
Much of this is unknown and uncertain. But many people are worried that the future may arrive sooner than expected. Repeatedly saying how difficult it would be to program a poker-playing bot (either GTO or not) is naive and seems to be oblivious to significant advances in this general area over the last few years.
To boil it down to one sentence, I am not concerned that Cepheus and Libratus will be seated at an online poker table any time soon, but the recent advances in AI techniques that led to the development of Cepheus and Libratus will undoubtedly lead to the advancement of many other winning online poker-playing bots.
Last edited by whosnext; 02-11-2017 at 05:39 AM.