Skip to content
This repository has been archived by the owner on Jul 16, 2024. It is now read-only.

Real time search #116

Open
CloneHub94 opened this issue Sep 29, 2020 · 3 comments
Open

Real time search #116

CloneHub94 opened this issue Sep 29, 2020 · 3 comments

Comments

@CloneHub94
Copy link

Just a few questions. I read the paper on Pluribus and there is one thing I don't seem to understand. The blueprint strategy used I assume IR 169-200-200-200 KE-KO buckets for its abstraction according to the supplementary material for Pluribus. But only gets used in the first round if the bet size is less then 100$ removed from any action abstraction in the blueprint strategy. However realtime search uses 500 buckets for future rounds and lossless abstraction in the round it is in.

  1. Does this mean that the algorithm uses lossless abstraction first and then buckets information situations to 200 buckets if it's for example on the flop(round 2) and then uses 500 buckets for rounds 3 and 4 in the subgame it is in?

  2. Given the fact that the supplementary material states that the algorithm only uses lossless abstraction in the first betting round but real-time search uses lossless abstraction in the round it is in and always gets used in rounds 2-4 could you explain how and when exactly lossless abstraction gets used?

I can't seem to understand how info sets are bucketed in this algorithm, could you explain?

@brewer-b
Copy link

Search is either done to a depth limit or to the end of the game. In either case, the current round uses lossless abstraction. If you are solving to a depth limit, there is no need for abstraction on future streets, as only the current street is part of the subgame. Headsup pots on the flop are solved to the end of the game using 500 buckets for the turn and river (these buckets are created separately for each flop).

Lossless abstraction is used in the first betting round in the blueprint because there are only 169 unique preflop hands.

@CloneHub94
Copy link
Author

CloneHub94 commented Aug 22, 2021

Thanks for your response.Just a small follow up question if you don't mind. How does the algorithm determine whether to use a depth-limit and how deep or to search to the end of the game?

Is there a metric that it uses to determine the depth of the depth-limit search?

@brewer-b
Copy link

brewer-b commented Aug 22, 2021

From the paper -
"If search is being done on the first betting round, then the subgame extends to the end of
the round, with leaf nodes at the chance nodes at the start of the second round. If search is
being done on the second betting round and there were more than two players at the start of the
round, then leaf nodes occur either at the chance nodes at the stat of the third betting round or
immediately after the second raise action, whichever is earlier. In all other cases, the subgame
extends to the end of the game"

Search is done on the first round if an opponent chooses a raise size that is more than $100 off from any raise
size in the blueprint action abstraction and there are no more than four players remaining in the hand.

Search is always done on rounds 2-4.

The decision comes down to if they have enough computing power to solve to the end of the game. This was what they decided on.

Sign up for free to subscribe to this conversation on GitHub. Already have an account? Sign in.
Labels
None yet
Projects
None yet
Development

No branches or pull requests

2 participants