Modulus magic!
Suppose we have `total_information` choices, and we first use them to encode the answer `x` to a question with `m` answers. That answer is encoded by the choice we take modulo `m`. How much "information" do we have left? That depends on the number of numbers less than `total_information` that are equal to `x` modulo `m`. Depending on the value of `x`, this is either `floor(total_information/m)` or `floor(total_information/m) + 1`. We now use all of this information as opposed to just `floor(total_information/m)`, at the cost of making our math not a lot more complicated but pretty confusing.
This commit is contained in:
parent
8ed01d47ca
commit
051ac7a097
2 changed files with 40 additions and 25 deletions
|
@ -73,5 +73,5 @@ On the first 20000 seeds, we have these average scores and win rates:
|
|||
|---------|---------|---------|---------|---------|
|
||||
| cheat | 24.8594 | 24.9785 | 24.9720 | 24.9557 |
|
||||
| | 90.59 % | 98.17 % | 97.76 % | 96.42 % |
|
||||
| info | 22.3427 | 24.7488 | 24.9055 | 24.9051 |
|
||||
| | 10.23 % | 81.67 % | 92.53 % | 92.62 % |
|
||||
| info | 22.3736 | 24.7840 | 24.9261 | 24.9160 |
|
||||
| | 10.41 % | 84.14 % | 94.33 % | 93.49 % |
|
||||
|
|
|
@ -19,20 +19,37 @@ impl ModulusInformation {
|
|||
Self::new(1, 0)
|
||||
}
|
||||
|
||||
pub fn combine(&mut self, other: Self) {
|
||||
pub fn combine(&mut self, other: Self, max_modulus: u32) {
|
||||
assert!(other.modulus <= self.info_remaining(max_modulus));
|
||||
self.value = self.value + self.modulus * other.value;
|
||||
self.modulus = self.modulus * other.modulus;
|
||||
self.modulus = std::cmp::min(max_modulus, self.modulus * other.modulus);
|
||||
assert!(self.value < self.modulus);
|
||||
}
|
||||
|
||||
pub fn info_remaining(&self, max_modulus: u32) -> u32 {
|
||||
// We want to find the largest number `result` such that
|
||||
// `self.combine(other, max_modulus)` works whenever `other.modulus == result`.
|
||||
// `other.value` can be up to `result - 1`, so calling combine could increase our value to
|
||||
// up to `self.value + self.modulus * (result - 1)`, which must always be less than
|
||||
// `max_modulus`.
|
||||
// Therefore, we compute the largest number `result` such that
|
||||
// `self.value + self.modulus * (result - 1) < max_modulus`.
|
||||
let result = (max_modulus - self.value - 1) / self.modulus + 1;
|
||||
assert!(self.value + self.modulus * (result - 1) < max_modulus);
|
||||
assert!(self.value + self.modulus * ((result + 1) - 1) >= max_modulus);
|
||||
result
|
||||
}
|
||||
|
||||
pub fn split(&mut self, modulus: u32) -> Self {
|
||||
assert!(self.modulus >= modulus);
|
||||
assert!(self.modulus % modulus == 0);
|
||||
let original_modulus = self.modulus;
|
||||
let original_value = self.value;
|
||||
self.modulus = self.modulus / modulus;
|
||||
let value = self.value % modulus;
|
||||
self.value = self.value / modulus;
|
||||
assert!(original_modulus == modulus * self.modulus);
|
||||
// `self.modulus` is the largest number such that
|
||||
// `value + (self.modulus - 1) * modulus < original_modulus`.
|
||||
// TODO: find an explanation of why this makes everything work out
|
||||
self.modulus = (original_modulus - value - 1) / modulus + 1;
|
||||
assert!(original_value == value + modulus * self.value);
|
||||
Self::new(modulus, value)
|
||||
}
|
||||
|
@ -42,11 +59,11 @@ impl ModulusInformation {
|
|||
self.modulus = modulus;
|
||||
}
|
||||
|
||||
pub fn cast_down(&mut self, modulus: u32) {
|
||||
assert!(self.modulus >= modulus);
|
||||
assert!(self.value < modulus);
|
||||
self.modulus = modulus;
|
||||
}
|
||||
// pub fn cast_down(&mut self, modulus: u32) {
|
||||
// assert!(self.modulus >= modulus);
|
||||
// assert!(self.value < modulus);
|
||||
// self.modulus = modulus;
|
||||
// }
|
||||
|
||||
pub fn add(&mut self, other: &Self) {
|
||||
assert!(self.modulus == other.modulus);
|
||||
|
@ -127,10 +144,10 @@ pub trait PublicInformation: Clone {
|
|||
let mut answer_info = ModulusInformation::none();
|
||||
{
|
||||
let callback = |hand_info: &mut HandInfo<CardPossibilityTable>, info_remaining: &mut u32, question: Box<Question>| {
|
||||
*info_remaining = *info_remaining / question.info_amount();
|
||||
let new_answer_info = question.answer_info(view.get_hand(player), view.get_board());
|
||||
question.acknowledge_answer_info(new_answer_info.clone(), hand_info, view.get_board());
|
||||
answer_info.combine(new_answer_info);
|
||||
answer_info.combine(new_answer_info, total_info);
|
||||
*info_remaining = answer_info.info_remaining(total_info);
|
||||
};
|
||||
self.ask_questions(player, hand_info, callback, total_info);
|
||||
}
|
||||
|
@ -146,18 +163,16 @@ pub trait PublicInformation: Clone {
|
|||
mut info: ModulusInformation,
|
||||
) {
|
||||
let total_info = info.modulus;
|
||||
{
|
||||
let callback = |hand_info: &mut HandInfo<CardPossibilityTable>, info_remaining: &mut u32, question: Box<Question>| {
|
||||
let q_info_amount = question.info_amount();
|
||||
*info_remaining = *info_remaining / q_info_amount;
|
||||
let info_modulus = info.modulus;
|
||||
// Instead of casting down the ModulusInformation to the product of all the questions before
|
||||
// answering any, we now have to cast down the ModulusInformation question-by-question.
|
||||
info.cast_down((info_modulus / q_info_amount) * q_info_amount);
|
||||
let answer_info = info.split(question.info_amount());
|
||||
question.acknowledge_answer_info(answer_info, hand_info, board);
|
||||
*info_remaining = info.modulus;
|
||||
};
|
||||
self.ask_questions(player, hand_info, callback, total_info);
|
||||
}
|
||||
assert!(info.value == 0);
|
||||
}
|
||||
|
||||
/// When deciding on a move, if we can choose between `total_info` choices,
|
||||
/// `self.get_hat_sum(total_info, view)` tells us which choice to take, and at the same time
|
||||
|
|
Loading…
Reference in a new issue