[davehcker]

I have posted this here before- hexafarms.com. I am trying to use ML to discover optimal phenotype for growing plants in vertical indoor farms to a. have the higest quality produce b. to lower the cost of producing leafy green/med plants, etc. within cities itself.

Basically, every leafy green (and herbs, and even mushrooms), can grow in a range of climatic condition (phenotype, roughly) ie temperature, humidity, water, CO2 level, pH, light (spectrum, duration and intensity) etc. As you might have seen around the world there is a rise in indoor vertical farms, but the truth is that 50% of those are not even profitable. My startup wants to discover the optimal parameters for each plant grown in our indoor vertical farm and eventually I would let our AI system control everything (something like alphaGo, but for growing plant X (lettuce, kale, chard, ). Think of it as reinforcement learning with live plants! I am betting on the fact that our startup will discover the 'plant recipes' and figure out the optimal parameters for the produce that we would grow. Then, the goal is that cities can grow food cheaper in more secure and sustainable way than our 'outsourced' approach in country side or far away lands.

So now I have secured some funding to be able to start working on optimizations, but I realized that *hardware* startups are such a different kind of beast (I am a good software product dev though, I think). Honestly, if anyone with experience in hardware related startups (or experience in the kind of venture I am in) would just want to meet me and advise me, I would take it any day. Being the star of the show, it's hard for me to handle market segmentation, tech dev, team, next round of funding, European tech landscape, etc. I am foreseeing so many ways that our decisions can kill my startup, all I need is advise from someone qualified/experienced enough. My email: david[at]hexafarms.com

[spiralganglion]

Reminder to focus on nutritive content, flavour, and crop diversity, not just yield. The past 100 years of industrial scale agriculture, with the singular goal of maximizing yields, has done incredible harm. (This has come up on HN repeatedly, so I trust you've seen it, but it's worth championing)

[dehrmann]

> incredible harm

I agree that micronutrient content has decreased in the past century. Some might be because of scale, some might be that yield gains are mostly driven by macronutrients and water, not micronutrients, it could be selecting varieties that taste better, or it could be depleting the soil.

That said, the US has an obesity epidemic, so there's no shortage of macronutrients. Macronutrient shortages also seem rare. Scurvy and rickets aren't exactly problems.

[apinstein]

This isn’t an answer to your ML question, but it is an answer to your problem.

I heard about a greenhouse company that has programmed their climate control to match “best growing conditions historical weather”. So, they ask local experts what year / location had the best X and then they use that region’s historical weather and replay it in their greenhouse. I thought that was brilliant!

(Just realized this was Kimbal Musk that mentioned this)

[eitland]

When I studied farming back in 1998-1999 we once visited a greenhouse and one interesting thing I picked up was that by observation some gardeners had realized that lowering the temperature a bit extra an hour or two before sunrise they could get their flowers to be more compact instead of stretching.

This had replaced shortening hormones in modern gardening (or at least at that greenhouse, but my understanding they were just doing the same thing as everyone else).

I guess there is a lot more to learn for those who have scale enough to experiment and patience to follow through.

[universa1]

Hmm,

Sounds similar to what I read a long time ago about a big tomato farm in the Netherlands... Have you tried talking to actual farmers of that produce? Universities? Agricultural faculties do a lot of research in that direction.

Expensive, quickly perishable produce might be able to compete, otherwise I guess free water and energy from above in the "remote" classical farming will be hard to beat.

And then my naive guess would be that to generate enough data for a "ml" approach not only by name might be somewhat expensive.

This sounds so negative, but this is not my intention... I wish you all the best and hopefully will stumble upon a success story in the future :-)

[3pt14159]

I know this isn't going to sound as sexy as AlphaGo for plants, but I really think this is a classic multilinear optimization problem once you've properly labeled the data and defined the dynamics between the plants / other organisms (e.g., aquaponics). You're looking to optimize multiple variables across a set of known constraints and I think if you properly defined these constraints you could save a lot of headache / buildout by leveraging a pre-exsting toolset like Excel with the Excel Solver add-in an a couple hundred user defined functions. We're talking 1% of the work to get something useable and product-market-fitable with automatic output of graphs, etc, that clients could tune and play with locally without you needing to actually share the source sauce. Eventually you could switch to Python for something more dynamic / web based.

[flir]

Yeah, the description made me think "simulated annealing" not "AI". I mean, even genetic algorithms might be overkill here.

[mattmanser]

I'm not able to help, but you don't have any contact details listed against your profile or in this post. How is anyone able to contact you?

At the very least what's a link to your startup's website?

[davehcker]

Sorry I thought my email was on my HN profile. I am sitting behind david[at]hexafarms.com

[sokoloff]

If you’ve listed it in the email field, that’s accessible to HN admins, but not users.

If you want users to have it from your profile, put it in the “about” field.

[mattmanser]

It sounds like an interesting project, good luck and I hope someone reaches out!

[mkolodny]

There's some great research on using evolutionary computation to explore plant growing recipes (light strength, how long to leave the lights on, etc). In one experiment, researchers discovered that basil doesn't need to sleep - it grows best with 24 hours of light per day. Risto Miikkulainen shared the experiment on Lex Fridman's podcast: https://youtu.be/CY_LEa9xQtg?t=27m7s I believe this is the paper describing that experiment: https://journals.plos.org/plosone/article?id=10.1371/journal...

[bravura]

This sort of ml problem is characterized by relatively expensive data labeling. Hence, hiring an expert or mixture of experts, and modeling the crop responses to their choices, will save you a lot of hill climbing The wrong part of the decision space

[exdsq]

That sounds awesome. I’d love to work in this field. Any tips on where you learn this stuff? Currently a software dev in crypto.

[qPM9l3XJrF]

I think you'd be better off using a Gaussian process than reinforcement learning

[ampdepolymerase]

You need to sequence the plants otherwise you will waste too much time on tuning hyperparameters.