There is much to dislike about this notion of "goodness" of a forecaster. As I said previously:

Despite the natural way to pose the concept of calibration, it is actually quite a weak notion of "goodness" of a forecaster. Let's take the weather example again, and imagine a weather forecaster who, despite the available data and the meteorological conditions, always predicts the base probability of rain. If the forecaster happens to know in advance that it rains about 13.2% of the time, then to remain calibrated she might as well predict this value every day. Hence, a well-calibrated but entirely useless forecaster.

Since being here in Israel I've had a change of heart -- Shie taught me a trick about calibration that rekindled my interest. He says it's "well-known" but it is new to me and so I'd like to share it.

The trick is this: you can use calibration as an optimal way to play repeated zero-sum games. Let's imagine we have a zero sum game with payoff matrix $M \in \mathbb{R}^{n \times m}$, where on each round $t$ player 1 choses a distribution $x_t \in \Delta_n$ over his actions, and player 2 then chooses an action $y_t \in \Delta_m$ (we can actually assume p2 plays deterministically). If the game is played repeatedly then the average payoff after $T$ rounds is $\frac 1 T \sum_{t=1}^T x_t M y_t$.

So let's assume player 1 has a calibrated forecasting algorithm for "guessing" p2's strategy, so let $y^?_t$ be the (randomized) guess of p2's action $y_t$. Then player 1 can "best respond" to this guess, that is by letting $x_t := \arg\max_{x \in \Delta_n} x M y^?_t$. Then under the calibrated forecasting assumption, this strategy is "optimal" in the limit, in that it earns at least the value of this zero-sum game. Why? While p1 might be guessing $y_t$ poorly, we know that because he is calibrated he's right "on average". So while he might be best responding to a bad forecast on each round, he is providing the best response "on average".

They key piece here, of course, is that we can achieve calibrated forecasting even when the outcome sequence is chosen by an adversary. For more details, see this paper by Sham Kakade and Dean Foster.

You're leaving us hanging here. Is the trick to use zero-sum games (against nature) for calibrated forecast?

ReplyDeletevery nice!

ReplyDeleteThis is nice! The payoff matrix can be used to represent how much correct guesses are worth (and incorrect ones also). So for example, it could be worth more to guess correctly that it is going to snow than to guess correctly that it is going to rain.

ReplyDeleteVery helpful website. Many important update information is here

ReplyDeleteThanks

cara menghilangkan kutil pada penis

ReplyDeletecara menghilangkan kutil di kemaluan pria

cara menghilangkan kutil di penis

cara mengobati kutil pada kemaluan pria

cara mengobati kutil pada penis

obat kutil di kemaluan pria

obat kutil pada kemaluan pria

obat kutil di alat kelamin pria

obat kutil kelamin pria

Obat kutil di kelamin pria

pengobatan kutil kelamin pada wanita

pengobatan kutil kelamin pada wanita hamil

obat kutil kelamin untuk wanita

obat kutil di kemaluan wanita

Cara menghilangkan kutil di sekitar kemaluan

cara menghilangkan kutil di sekitar kemaluan

cara menghilangkan kutil di daerah kemaluan

cara menghilangkan kutil di dekat kemaluan

obat untuk kutil kelamin

obat kutil kelamin di apotik

obat kutil kelamin tradisional

obat kutil kelamin pria

obat kutil kelamin wanita