Echoing what everyone else is saying - I have no idea what this does. Maybe I’m not in the target market and I’d jump on it if I were, but real world examples would help.
And I don’t mean real world examples like you’ve listed. Those are just names of domains (neurofeedback, order fulfillment). I can list domains too (accounting, genomics). Give me a case study of what your thing does, with the real world payoff.
Here’s an example (trying to guess what it does, could be way off):
Imagine you had a time series of the temperature in your room every day and when your AC engages
If you had a time series ML engine, it could optimize when the AC turns on and off
This would reduce your energy usage by not overcooling the room at the end of the day as the temp drops naturally
See how that format works? Situation without your thing. What your thing can do. Real world benefit user gets from using your thing.
Thanks for taking the time to write up this feedback. Yes it's obvious now we need to improve the examples, and I like your structure. Perhaps you can help us.
I'm having trouble understanding what the goal is of this as well. It seems like the quick summary would be "ML-based forecasting/prediction in a box" but the readme is making all of these broad claims...
We're looking to make timeseries AI easier for developers to integrate into their applications by providing tooling and patterns that are familiar to them. We're a group of devs that were looking to add intelligence to one of our projects, but struggling with existing tooling. We wanted to use patterns that were more familiar to us as devs, like a quick debugging loop, easily consumed packages, etc.
Note: has nothing to do with the long-standing ecosystem of circuit simulators, as far as I can tell. Although, hey, try pointing it at some transients, maybe we can have AIs design circuits, that could be nice.
Hello Gimpei. Spice.ai currently supports two algorithms, Vanilla Policy Gradient and Deep Q-Learning. It provides an interface to plug in your own algorithms, though. We're looking to add more to it as we go along.
I think you are going to confuse people by framing this as "time series" and then focussing on reinforcement learning. A lot/most of your framing is around RL.
from a first glance and a read through your roadmap, this does not feel suitable for people who know what they are doing with RL. It also does not feel suitable for people who don't know what they are doing with RL.
We chose them because they were fairly straightforward to implement and different enough from one another that we could ensure our interface generalized well.
Re benchmarking - at this point we're looking to show directionality, not necessarily blinding speed. We intend to get the tooling feeling right, then work to optimize perf.
Right now, training data comes from the local disk, InfluxDB, or can be piped in from your application via our API. We're looking to build out a set of community-driven components for streaming and processing data. You can learn more about that here - https://github.com/spiceai/data-components-contrib
So I am not sure that I understand you. Why would you even implement a forecasting algorithm and not use open source one. (Unless I am missing something).
Also, how do you plan to verify that the algorithm works?
Note, that your customers would need to make critical business decisions based on this software, so I would refrain doing clean room impl of the forecasting algorithm.
You seem to be using Go as well as Python for your project. Are you calling models written in python using Go?
It's rare to see Go used in ML projects(Perhaps lack of batteries like Numpy,Pandas etc.), Which is a shame because I think Go is a perfect replacement for Python and it helps to build production ready ML applications off the bat without the performance limitations of Python.
Thank you! Yes, that's what we are currently doing, however the long term direction is to move as much of the codebase to Go/Rust as possible to be portable across hosting environments as possible.
How does it perform with live data?
Can I feed a continuous stream of data and it recommends the next action?
Can I limit the dataset that is taken into account (look-back amount)?
I know this may be a bit pedantic, and from a marketing perspective I'm sure everyone is telling you that "you have to market it as AI!", but my pet peeves is that the phrase "time series AI" doesn't make any sense in English. "ML for time series data" makes much more sense to me, is valid English, and from your post sounds like what you're actually doing.
Pfft that's so 2019, it's gotta be IoT cloud edge AI these days.
...because what really matters to most software devs is apparently increasing their job security by pushing processing and energy requirements onto the customer while roping them into a monthly support contract so they can be milked indefinitely.
Thanks for taking the time to write up this feedback. Great point. We've been coding for three months, and are still figuring out exactly how to message the project.
> Try:
> ServerOps sample - a more in-depth version of the quickstart you just completed, using CPU metrics from your own machine
> Gardener - Intelligently water a simulated garden
> Trader - a basic Bitcoin trading bot
- https://github.com/spiceai/samples/tree/trunk/serverops/READ...
- https://github.com/spiceai/samples/tree/trunk/gardener/READM...
- https://github.com/spiceai/quickstarts/tree/trunk/trader/REA...