Set mean reversion target to D0(4)

[user1823]

Supersedes and closes #123

Coding it like this will allow tweaking the formula further without needing to change the code at two places.

[user1823]

By the way, shouldn't the simulator import the definitions from fsrs_optimizer.py instead of defining them again?

fsrs-optimizer/src/fsrs_optimizer/fsrs_simulator.py

Lines 80 to 132 in 86eeb69

	def stability_after_success(s, r, d, rating):
	hard_penalty = np.where(rating == 2, w[15], 1)
	easy_bonus = np.where(rating == 4, w[16], 1)
	return np.maximum(
	0.01,
	s
	* (
	1
	+ np.exp(w[8])
	* (11 - d)
	* np.power(s, -w[9])
	* (np.exp((1 - r) * w[10]) - 1)
	* hard_penalty
	* easy_bonus
	),
	)
	def stability_after_failure(s, r, d):
	return np.maximum(
	0.01,
	np.minimum(
	w[11]
	* np.power(d, -w[12])
	* (np.power(s + 1, w[13]) - 1)
	* np.exp((1 - r) * w[14]),
	s,
	),
	)
	def stability_short_term(s, init_rating=None):
	if init_rating is not None:
	rating_offset = np.choose(init_rating - 1, first_rating_offset)
	session_len = np.choose(init_rating - 1, first_session_len)
	else:
	rating_offset = forget_rating_offset
	session_len = forget_session_len
	new_s = s * np.exp(w[17] * (rating_offset + session_len * w[18]))
	return new_s
	def init_d(rating):
	new_d = w[4] - np.exp(w[5] * (rating - 1)) + 1
	rating_offset = np.choose(rating - 1, first_rating_offset)
	new_d -= w[6] * rating_offset
	return np.clip(new_d, 1, 10)
	def next_d(d, rating):
	new_d = d - w[6] * (rating - 3)
	new_d = mean_reversion(w[4], new_d)
	return np.clip(new_d, 1, 10)
	def mean_reversion(init, current):
	return w[7] * init + (1 - w[7]) * current

[L-M-Sherlock]

Due to type inference problem, it's hard to reuse this function between optimizer and simulator.

By the way, you forget to call the member function as self.init_d.

[user1823]

By the way, you forget to call the member function as self.init_d.

Thanks. If you don't mind, can you briefly explain what is the difference between self.init_d and init_d?

[user1823]

Due to type inference problem, it's hard to reuse this function between optimizer and simulator.

Then, should I make the corrections in the simulator too?

[L-M-Sherlock]

Thanks. If you don't mind, can you briefly explain what is the difference between self.init_d and init_d?

The init_d function is defined under the class FSRS, so it's the member method. It's a concept from object-oriented programming.

Then, should I make the corrections in the simulator too?

Sure, please. I will check it later.

[user1823]

I have edited the simulator too.

But, I don't really understand what is rating_offset.

fsrs-optimizer/src/fsrs_optimizer/fsrs_simulator.py

Lines 119 to 123 in 40376ed

	def init_d(rating):
	new_d = w[4] - np.exp(w[5] * (rating - 1)) + 1
	rating_offset = np.choose(rating - 1, first_rating_offset)
	new_d -= w[6] * rating_offset
	return np.clip(new_d, 1, 10)

Also, there is one inconsistency. In the optimizer, mean reversion is using the unclamped value of D0(4). But, in the simulator, mean reversion is using the clamped value of D0(4). Which one do you think is more appropriate?

[L-M-Sherlock]

I make them consistent in the last commit. rating_offset is hard to explain. In short, it's used to simulate the effect of short-term reviews.

[Expertium]

Rating offset is basically "average grade of same-day reviews, minus 3".
For example, if you pressed Again and Hard, that's 1 and 2. (1+2)/2 = 1.5. So the average is 1.5. 1.5-3=-1.5. -1.5 is the offset.
EDIT: I'm confused, seems like in this case it's not minus 3, but minus 1 instead. Well, hopefully LMSherlock can explain it better.

[user1823]

OK, LGTM.

Note that this PR doesn't include updates to the default parameters from #123.

[L-M-Sherlock]

EDIT: I'm confused, seems like in this case it's not minus 3, but minus 1 instead. Well, hopefully LMSherlock can explain it better.

Any problem?

[user1823]

I just realized that this is not a bug fix and you deliberately used D0(1) for mean reversion previously (Found out by reading the Wiki page).

So, have you tried using D0(3) to see if it gives better results?

[Expertium]

We recently tried D0(1) vs D0(3) vs D0(4), and they all give almost the same results. RMSE differs in the 4 digit after the decimal point. So LMSherlock decided to go with D0(4)

[L-M-Sherlock]

LGTM