Hacker News new | past | comments | ask | show | jobs | submit login

We tried to make the paper as accessible as possible. A lot of these questions are covered in the supplementary material (along with pseudocode).

- Are the action and information abstraction procedures hand-engineered or learned in some manner?

- How does it decide how many bets to consider in a particular situation?

The information abstraction is determined by k-means clustering on certain features. There wasn't much thought put into the action abstraction because it turns out the exact sizes you use don't matter that much as long as the bot has enough options to choose from. We basically just did 0.25x pot, 0.5x pot, 1x pot, etc. The number of sizes varied depending on the situation.

- Is there anything interesting going on with how the strategy is compressed in memory?


- How do you decide in the first betting round if a bet is far enough off-tree that online search is needed?

We set a threshold at $100.

- When searching beyond leaf nodes, how did you choose how far to bias the strategies toward calling, raising, and folding?

In each case, we multiplied by the biased action's probability by a factor of 5 and renormalized. In theory it doesn't really matter what the factor is.

- After it calculates how it would act with every possible hand, how does it use that to balance its strategy while taking into account the hand it is actually holding?

This comes out naturally from our use of Linear Counterfactual Regret Minimization in the search space. It's covered in more detail in the supplementary material

- In general, how much do these kind of engineering details and hyperparameters matter to your results and to the efficiency of training? How much time did you spend on this? Roughly how many lines of code are important for making this work?

I think it's all pretty robust to the choice of parameters, but we didn't do extensive testing to see. While these bots are quite easy to train, the variance is so high in poker that getting meaningful experimental results is relatively quite computationally expensive.

- Why does this training method work so well on CPUs vs GPUs? Do you think there are any lessons here that might improve training efficiency for 2-player perfect-information systems such as AlphaZero?

I think the key is that the search algorithm is picking up so much of the slack that we don't really need to train an amazing precomputed strategy. If we weren't using search, it would probably be infeasible to generate a strong 6-player poker AI. Search was also critical for previous AI benchmark victories like chess and Go.

Guidelines | FAQ | Support | API | Security | Lists | Bookmarklet | Legal | Apply to YC | Contact