What happened to RAI? It seemed very promising. The price has actually been reasonably stable-ish for quite a while (since a rocky start).

We need a properly decentralized algo stable more than ever right now.

Pragmatism (Security) at OP Labs.
Prev: ConsenSys & Coinbase.

Attribs: Friendly, curious, introverted, mid-witted.
Chans: /eth-security /maurelian /stoicism




There's a bunch of analysis posts that have been written over the years, I can dig them up. IIRC the tldr is the price was not meant to be pegged, but rather dampened against volatility of the collateral (ETH). It used positive/negative interest rates as incentives to push/dampen the price action. But that has a cost, so I think an equalibrium of volatility around an average price point trends downwards slowly (as that "dampening" cost is slowly extracted).

HAI is a similar design but introduces multiple collateral assets with yield earning, so it should do a better job maintaining a more stable price target despite the dampening cost.

This is the best explainer I've found if that's what you're thinking of.
https://dankradfeist.de/ethereum/2023/01/31/rai-crypto-experiment.html

I have a decent understanding of the mechanism (classic PID controller), and that you're basically getting dampened but not zero volatility in exchange for stronger decentralization. But I can see how the consistent downtrend in '23 could turn off a lot of people, and the local spikes are pretty big which I see as more of a negative than the longer term volatility

A doodler and computerer. I like permissive/permissionless open source, smart contracts, p2p systems, room-scale VR, and NixOS.

shazow.net

Currently: WhatsABI

PID controllers are very flexible controllers, but the flip side of tuning them to maintain stability in a very dynamic system is that you just have to be super conservative about response. Hence why Rai has like 1 month response where it closes down on volitlity. If you made that like 1 hour it would be completely unstable

I have been waiting to see more advanced designs using results from modern control theory (which has been around since the 60s) as you could probably do something really cool with more advanced model weights, feed-forward mechanisms, etc. But... Nothing of that sort yet.

God I wish I had time to develop these ideas myself

Oh, what is the 1 month parameter is that how long it takes to adjust the PID params?

The integral gain (I in PID), they call it "redemption rate" I think

In a classical PID controller, the integral gain is meant to slowly adjust for "error drift" where the faster Proportional and Derivative gains cause an adjustment that doesn't quite match the target signal

The main thing to know about PID controllers is that they're designed to work for linear, time-invariant systems. Linear as in the dynamic system does not have a complex to external stimulus, and time-invariant as in it does not modify its response over time. Crypto money markets are neither.

Even for a propeller in a relatively stable medium like water, there are chaotic motions in the medium that would be very difficult to model well as LTI. But that's what makes PID powerful, is you can just simplify the system into looking LTI *most of the time* then you can tune PID to work for it quite easily. You can do that for a propeller, because except for the corner cases of starting from a dead stop (which is kind of insignificant), most of the time it is pretty stable flow, and you have a human operator who is able to adjust the operation of the controller (by providing fuel inputs) to get what they are looking for (simple engine controllers typically do not have an I term in them for this reason)

Ahh, good context thanks.  
My intuition from uni is that you'd typically use it for something like a propeller or cruise control, but didn't consider how unsuited it would be for such a complex system.

I wonder if there are tradfi rebalancing mechanisms using better algos.