Options

Azimuth meeting in Sheffield

After a bit of struggle to get ourselves in the same place at the same time, we had a meeting of Azimuthers in Sheffield on Thursday 28 March. Glyn Adgie, Jim Stuttard, David Tweed, and myself in attendance.

It was fun meeting them, and as usual they looked completely different than the mental images I'd built up over the years. I learned a lot more about what they do with computers in their work and side-projects. I wish they'd talk more about that stuff here, because it's very interesting, yet also sufficiently technical that I have trouble remembering it in perfect detail!

Let me try to remember the most important things we said:

  1. Jim and Glyn and I seemed to agree that it was good to continue developing simple climate models that could run online, either on the user's browser or on the server Jim is developing. It may not "save the planet" quite as much as I'd like. But it's something we can actually do, and it can educate lots of people. They've put more work into getting ready to do this than I'd realized.

  2. Jim wanted to charge through Gerald North's book on simple climate models and program those models up. He'd been held back somewhat by not having access to the figures in this book. This is something I should deal with. I should contact North and say we want to do this.

  3. Jim was also having trouble finding where he'd gotten a specific Budyko-Sellars model (that is, a simple energy balance model of the Earth's climate) from Nathan Urban. Can we figure this out?

  4. Jim and Glyn and David and I seemed to agree that it would be good to learn more about analyzing time series data for signs of incipient tipping points, or other forms of instability. This seems to score high on both planet-saving potential and doability. I'd have to learn a bunch more stuff, but it fits in with my new obsession with control theory. We might be able to bring David's skills at dealing with "big data sets" into play. Also, this could interact well with Gloria Gonzalez-Rivera's interest in systemic risk in financial markets, and its relation to the stability of food webs - she wants to start a seminar on that next year, and I should remind her that this is very interesting to me.

  5. Jim and I seemed to agree that it would be fun and useful to have some software that'll make it easy for people (e.g., me and my grad students) to investigate the behavior of stochastic Petri nets. In retrospect this seems like a lower priority than item 1, at least until I find some models that shed light on something important, e.g. the stability of food webs. But Jim said he's already scoured the planet for Petri net software. I really want a list of the software he found, with some comments about it! I want this sort of information to be the last chapter of my book with Jacob Biamonte.

  6. David Tweed said he would like to work on mitigation of global warming, since ultimately understanding a problem is insufficient; we need to be doing something about it. He wasn't quite sure what would be best to do. I agree completely with all this. He suggested thinking about 'smart grids' that can deal with intermittent power sources like wind and solar. That fits in with the 'network theory' theme, and also with control theory.

  7. I urged Jim, but really all three of them, to write some short blog posts about what they're thinking about. They are thinking about a lot of interesting things, and I fear that my role as the 'public face of Azimuth' limits the number of interesting things that get talked about. Of course I'm not suggesting that I post less stuff; rather that they post more.

  8. Jim had noticed a feature on the Azimuth Forum that I'd never seen: you can get it to notify you when new comments appear. However, it seemed to work only intermittently... maybe. I told him to talk to Andrew Stacey about this. He should post a comment in the 'Technical' category about this.

This is what I remember right now. Jim, Glyn and David will note that I've left out all their more technical comments on software - maybe they can fix that deficit.

We should talk more often!

Comments

  • 1.
    edited April 2013

    Sounds great.

    About 3) Do you mean this comment by Nathan Urban?

    I shoudl remind myself to read about this too, to see how much it aligns with the logarithmic forcing (I remember some notes where some comment is made about simple climate models being sometimes too simple, but this was just a vague general remark, not specifically aimed at Budyko-Sellers, but I should check)

    Comment Source:Sounds great. About 3) Do you mean this [comment by Nathan Urban](http://forum.azimuthproject.org/discussion/617/qday-to-mu/?Focus=3816#Comment_3816)? I shoudl remind myself to read about this too, to see how much it aligns with the logarithmic forcing (I remember some notes where some comment is made about simple climate models being sometimes too simple, but this was just a vague general remark, not specifically aimed at Budyko-Sellers, but I should check)
  • 2.
    edited April 2013

    Hi Frederick,

    Glyn and I had a great time in Sheffield and met some other really interesting people apart from David and John. I got a quite long term question of mine answered which means I'll have a go at learning Agda and not Epigram or Coq thanks to Neil Ghani.

    It's not that comment. It's this code which I apparently posted a year ago but had forgotten. I got the parameter values from Nathan Urban but from where I don't know. Nathan might help me out again if he sees this post.

    testCO2 = temp 4.3
    
    temp f = (-f)/totalF where
    
       lambda0 = 3.2  :: Double -- Plank
    
       totalF   = avLambda - lambda0
    
       avLambda     = (maxSumF - minSumF)/2  
    
       minSumF = foldl (+) 0 (fmap fst lambda)
    
       maxSumF = foldl (+) 0 (fmap snd lambda)
    
       lambda = [waterVapour,lapseRate,clouds,albedo]
    
       waterVapour = (1.48,2.14)
    
       lapseRate = (-0.41,-1.27)
    
       clouds = (0.18,1.18)
    
       albedo = (0.07,0.34)
    

    from here.

    However thanks very much for the link as there's code there I never knew existed!

    I've also found existing pages of and about petri net software so i should add a few notes somewhere there.

    I didn't remember that Dave Tanzer was working on new Python Petri net software which I've just tried and which runs fine :).

    Comment Source:Hi Frederick, Glyn and I had a great time in Sheffield and met some other really interesting people apart from David and John. I got a quite long term question of mine answered which means I'll have a go at learning Agda and not Epigram or Coq thanks to Neil Ghani. It's not that comment. It's this code which I apparently posted a year ago but had forgotten. I got the parameter values from [[Nathan Urban]] but from where I don't know. Nathan might help me out again if he sees this post. testCO2 = temp 4.3 temp f = (-f)/totalF where lambda0 = 3.2 :: Double -- Plank totalF = avLambda - lambda0 avLambda = (maxSumF - minSumF)/2 minSumF = foldl (+) 0 (fmap fst lambda) maxSumF = foldl (+) 0 (fmap snd lambda) lambda = [waterVapour,lapseRate,clouds,albedo] waterVapour = (1.48,2.14) lapseRate = (-0.41,-1.27) clouds = (0.18,1.18) albedo = (0.07,0.34) from [here](http://forum.azimuthproject.org/discussion/938/azimuth-code-project-questions/#Item_32). However thanks very much for the link as there's code there I never knew existed! I've also found existing pages of and about petri net software so i should add a few notes somewhere there. I didn't remember that [[Dave Tanzer]] was working on new Python Petri net software which I've just tried and which runs fine :).
  • 3.
    edited April 2013

    In 5), I asked Jim for a list of the software he'd found for working with Petri nets. If he likes, he can add that information to the wiki here:

    Just make sure to tell us if you do!

    There's already some information there, namely:


    There is, or was, a visual formalism called DEVS—Discrete Event System Specification—which has developed a host of visual tools for modeling and maybe code/stub gen (if I remember correctly. These are also other types of Open source packages on wikipedia:

    • CD++ is an open source based on the DEVS and Cell-DEVS formalisms (a discrete-event specification of cellular automata models). Hundreds of model samples are available. Varied 2D and 3D visualization engines can be used to improve the analysis of the simulation results.
    • PowerDEVS is an integrated tool for hybrid systems modeling and simulation based on the DEVS formalism.
    • SimPy is an open source process-oriented discrete event simulation package implemented in Python. It is based on Simula concepts, but goes significantly beyond Simula in its synchronization constructs.
    • Tortuga is an open source software framework for discrete-event simulation in Java.
    • Facsimile is a free, open-source discrete-event simulation/emulation library.
    • Galatea - Galatea is a Agent-based simulation platform.
    • MASON is a fast discrete-event multiagent simulation library core in Java, designed to be the foundation for large custom-purpose Java simulations.
    Comment Source:In 5), I asked Jim for a list of the software he'd found for working with Petri nets. If he likes, he can add that information to the wiki here: * [Petri net - software](http://www.azimuthproject.org/azimuth/show/Petri+net#Software). Just make sure to tell us if you do! There's already some information there, namely: <hr/> There is, or was, a visual formalism called DEVS&mdash;Discrete Event System Specification&mdash;which has developed a host of visual tools for modeling and maybe code/stub gen (if I remember correctly. These are also other types of [Open source packages on wikipedia](http://en.wikipedia.org/wiki/List_of_discrete_event_simulation_software): * CD++ is an open source based on the DEVS and Cell-DEVS formalisms (a discrete-event specification of cellular automata models). Hundreds of model samples are available. Varied 2D and 3D visualization engines can be used to improve the analysis of the simulation results. * PowerDEVS is an integrated tool for hybrid systems modeling and simulation based on the DEVS formalism. * SimPy is an open source process-oriented discrete event simulation package implemented in Python. It is based on Simula concepts, but goes significantly beyond Simula in its synchronization constructs. * Tortuga is an open source software framework for discrete-event simulation in Java. * Facsimile is a free, open-source discrete-event simulation/emulation library. * Galatea - Galatea is a Agent-based simulation platform. * MASON is a fast discrete-event multiagent simulation library core in Java, designed to be the foundation for large custom-purpose Java simulations.
  • 4.

    I've made a start on directly copying the first of my dnr;tl rough notes to the wiki. John, I'm sorry to mess up your heading styles a bit. Perhaps you could normalise some of our stuff? All your packages are in a list and mine are sub-sub... etc. heading fonts. I'll try and clean up what I've written, find, cleanup and post some other notes and add the missiing links in the next day or so.

    Comment Source:I've made a start on directly copying the first of my dnr;tl rough notes to the wiki. John, I'm sorry to mess up your heading styles a bit. Perhaps you could normalise some of our stuff? All your packages are in a list and mine are sub-sub... etc. heading fonts. I'll try and clean up what I've written, find, cleanup and post some other notes and add the missiing links in the next day or so.
  • 5.

    I don't know where the parameter values in that model came from (lapse rate, clouds, etc.). I don't think they were from me; I wouldn't know what to set them to without doing a literature search.

    About tipping points, control theory, etc.:

    I've talked to various people who have looked at the "detecting tipping points" problem using various methods (Bayesian prediction, looking at changes in the noise spectrum, etc.). They seem to feel that in most cases, you can't detect a tipping point very far in advance. Thus you probably wouldn't be able to avert it. Maybe you'd be able to start adapting earlier.

    I've maintained a long interest in control theory - mostly computationally, using approximate dynamic programming, rather than the more mathematical approaches like the Hamilton-Jacobi-Bellman equation. But I haven't done much with it. I hope to get into this more over the next year, but it will be probably closer to "in another year" than "soon". I'm interested in applications to climate mitigation (e.g. economics of emissions abatement ), and climate adaptation (e.g. coastal preparations for sea level). I'm also interested in active learning (where you're making decisions under uncertainty and can adapt your policies as you learn.) I know people here who do control theory for the smart grid. It might be fun to start with some toy environmental problem like managing a fishery described by the Lotka-Volterra equation.

    Comment Source:I don't know where the parameter values in that model came from (lapse rate, clouds, etc.). I don't think they were from me; I wouldn't know what to set them to without doing a literature search. About tipping points, control theory, etc.: I've talked to various people who have looked at the "detecting tipping points" problem using various methods (Bayesian prediction, looking at changes in the noise spectrum, etc.). They seem to feel that in most cases, you can't detect a tipping point very far in advance. Thus you probably wouldn't be able to avert it. Maybe you'd be able to start adapting earlier. I've maintained a long interest in control theory - mostly computationally, using approximate dynamic programming, rather than the more mathematical approaches like the Hamilton-Jacobi-Bellman equation. But I haven't done much with it. I hope to get into this more over the next year, but it will be probably closer to "in another year" than "soon". I'm interested in applications to climate mitigation (e.g. economics of emissions abatement ), and climate adaptation (e.g. coastal preparations for sea level). I'm also interested in active learning (where you're making decisions under uncertainty and can adapt your policies as you learn.) I know people here who do control theory for the smart grid. It might be fun to start with some toy environmental problem like managing a fishery described by the Lotka-Volterra equation.
  • 6.

    That's a mystery. I've tried searching on the Forum and the Wiki for lapse rate etc. to no avail. I couldn't have made them up, even if I' wanted to :). I'll carry on looking.

    it's bad news to read that tipping point detection isn't going to give us much warning but thanks for the guidance.

    Comment Source:That's a mystery. I've tried searching on the Forum and the Wiki for lapse rate etc. to no avail. I couldn't have made them up, even if I' wanted to :). I'll carry on looking. it's bad news to read that tipping point detection isn't going to give us much warning but thanks for the guidance.
  • 7.

    Wish I'd known that there was a planned meet-up ... would've been interesting to put some faces to names.

    Comment Source:Wish I'd known that there was a planned meet-up ... would've been interesting to put some faces to names.
  • 8.
    edited April 2013

    Hi, apologies for the slow response over the Easter break.

    Yes, the meet up in Sheffield was great and I'm only sorry that I had to rush off to catch my train. It was great to meet people, and it's just a pity that I'm currently so focussed on an end-of-April deadline I've got that I perhaps wasn't in the moment as much as I could have been. It was also a bit unfortunate I didn't realise Andrew was interested in putting faces to names: I actually thought to myself "That's Andrew Stacey" when I saw him leave the lecture theatre, but I didn't feel like accosting him on the off-chance...

    Anyway, just to touch on a couple of points that have come up: while I do think that it's probably unlikely that further analysis of changes in the climate and biosphere are to be listened to, I'm interested in looking at responses for a range of different reasons (EDIT: double negative removed). Partly it's because that's probably where my skills apply best (prototyping, initial data-analysis, etc) most apply but also because dealing with responses seems to be a bigger area and hence there's more scope for doing stuff that's not already being done. All that said, it's not completely clear what concrete things to work on... One of the things that most interests me is the possibilities for detecting patterns in mass behaviour.

    Regarding Nathan's interest in control theory, that sounds quite interesting. Although I'm probably even more interested in the "non-analytical" methods of control than even Nathan, I'm partly (because of my background) interested in models which are "condensed" versions of training data so I'm not sure how much of import about control is learnable using an analytical model (such as Lotka-Volterra) to generate the system to learn about. But I'm certainly interested in participating in discussions on the subject!

    Regarding John's point about writing blog posts that are as much about questions as answers, that's one possible area where I could contribute more than I have (although I'm still a little hesitant about throwing out commentary I can't back up with solid facts out there... :-) ) The other thing is that most of what I've been thinking about recently is some of the mathematical/algorithmic problems that come up in compiling mathermatical formulae, so not really environmentally focussed stuff.

    Anyway, hopefully I'll have some more "discretionary time" to spend on Azimuth type stuff soon. (Incidentally, I've recently learned that not only is 17 per cent of the web traffic made up of lol-cats and other animal pictures), it's even getting to the point where they're dominating "corrected searches": I tried a google with "cut DAG" in the terms and it decided to search for "cute dog", requiring me to explicitly say "yes I do actually want to search for my original terms".)

    Comment Source:Hi, apologies for the slow response over the Easter break. Yes, the meet up in Sheffield was great and I'm only sorry that I had to rush off to catch my train. It was great to meet people, and it's just a pity that I'm currently so focussed on an end-of-April deadline I've got that I perhaps wasn't in the moment as much as I could have been. It was also a bit unfortunate I didn't realise Andrew was interested in putting faces to names: I actually thought to myself "That's Andrew Stacey" when I saw him leave the lecture theatre, but I didn't feel like accosting him on the off-chance... Anyway, just to touch on a couple of points that have come up: while I do think that it's probably unlikely that further analysis of changes in the climate and biosphere are to be listened to, I'm interested in looking at responses for a range of different reasons (EDIT: double negative removed). Partly it's because that's probably where my skills apply best (prototyping, initial data-analysis, etc) most apply but also because dealing with responses seems to be a bigger area and hence there's more scope for doing stuff that's not already being done. All that said, it's not completely clear what concrete things to work on... One of the things that most interests me is the possibilities for detecting patterns in mass behaviour. Regarding Nathan's interest in control theory, that sounds quite interesting. Although I'm probably even more interested in the "non-analytical" methods of control than even Nathan, I'm partly (because of my background) interested in models which are "condensed" versions of training data so I'm not sure how much of import about control is learnable using an analytical model (such as Lotka-Volterra) to generate the system to learn about. But I'm certainly interested in participating in discussions on the subject! Regarding John's point about writing blog posts that are as much about questions as answers, that's one possible area where I could contribute more than I have (although I'm still a little hesitant about throwing out commentary I can't back up with solid facts out there... :-) ) The other thing is that most of what I've been thinking about recently is some of the mathematical/algorithmic problems that come up in compiling mathermatical formulae, so not really environmentally focussed stuff. Anyway, hopefully I'll have some more "discretionary time" to spend on Azimuth type stuff soon. (Incidentally, I've recently learned that not only is 17 per cent of the web traffic made up of lol-cats and other animal pictures), it's even getting to the point where they're dominating "corrected searches": I tried a google with "cut DAG" in the terms and it decided to search for "cute dog", requiring me to explicitly say "yes I do actually want to search for my original terms".)
  • 9.
    edited April 2013

    Dave: I think you might have made an optimistic Freudian typo:

    " I do think that it’s probably unlikely that further analysis of changes in the climate and biosphere are unlikley to be listened to,"

    I've just emailed you and Glyn as I failed to build Julia on the server.

    Comment Source:Dave: I think you might have made an optimistic Freudian typo: " I do think that it’s probably unlikely that further analysis of changes in the climate and biosphere are unlikley to be listened to," I've just emailed you and Glyn as I failed to build Julia on the server.
  • 10.

    Thanks for spotting the typo; I have a bad habit of editing sentences to improve the flow and not removing all the original bits.

    Comment Source:Thanks for spotting the typo; I have a bad habit of editing sentences to improve the flow and not removing all the original bits.
  • 11.

    Andrew wrote:

    Wish I’d known that there was a planned meet-up … would’ve been interesting to put some faces to names.

    Sorry! We'd been discussing the meetup here, and of course I also knew you'd be there, but for some reason I didn't think of telling you about this.

    Comment Source:Andrew wrote: > Wish I’d known that there was a planned meet-up … would’ve been interesting to put some faces to names. Sorry! We'd been discussing the meetup here, and of course I also knew you'd be there, but for some reason I didn't think of telling you about this.
  • 12.

    So, maybe it's less important to detect 'tipping points' than something like 'leverage points': things to do that would have a big effect. This might be especially true if we could model mass behavior.

    Nathan wrote:

    I’ve maintained a long interest in control theory - mostly computationally, using approximate dynamic programming, rather than the more mathematical approaches like the Hamilton-Jacobi-Bellman equation. But I haven’t done much with it. I hope to get into this more over the next year...

    What sort of project might you get into?

    David wrote:

    Although I’m probably even more interested in the “non-analytical” methods of control than even Nathan, I’m partly (because of my background) interested in models which are “condensed” versions of training data so I’m not sure how much of import about control is learnable using an analytical model (such as Lotka-Volterra) to generate the system to learn about.

    I get the feeling everyone thinks I'm committed to analytically solvable models. I guess the guilty conscience needs no direct accuser. But the way I see it, it's not analytically methods that interest me: I just know that whatever work I'll be doing will be mainly pencil-and-paper. Of course this includes "analytical methods". But it also includes "setting up frameworks", "developing rigorous analogies", "getting different communities to talk to each other", and other meta-useful tasks.

    Comment Source:So, maybe it's less important to detect 'tipping points' than something like 'leverage points': things to do that would have a big effect. This might be especially true if we could model mass behavior. Nathan wrote: > I’ve maintained a long interest in control theory - mostly computationally, using approximate dynamic programming, rather than the more mathematical approaches like the Hamilton-Jacobi-Bellman equation. But I haven’t done much with it. I hope to get into this more over the next year... What sort of project might you get into? David wrote: > Although I’m probably even more interested in the “non-analytical” methods of control than even Nathan, I’m partly (because of my background) interested in models which are “condensed” versions of training data so I’m not sure how much of import about control is learnable using an analytical model (such as Lotka-Volterra) to generate the system to learn about. I get the feeling everyone thinks I'm committed to analytically solvable models. I guess the guilty conscience needs no direct accuser. But the way I see it, it's not analytically methods that interest me: I just know that whatever work I'll be doing will be mainly pencil-and-paper. Of course this includes "analytical methods". But it also includes "setting up frameworks", "developing rigorous analogies", "getting different communities to talk to each other", and other meta-useful tasks. <img src = "http://math.ucr.edu/home/baez/emoticons/tongue2.gif" alt = ""/>
  • 13.

    John,

    So, maybe it’s less important to detect ’tipping points’ than something like ’leverage points’: things to do that would have a big effect. This might be especially true if we could model mass behavior.

    Yes, although these are perhaps harder to identify and depend more strongly on socioeconomic assumptions.

    What sort of project might you get into?

    I'm not sure yet. I have a few ideas of things to play around with.

    There has a bit of work on "endogenous learning" about climate feedbacks, i.e. how do mitigation policies change when you're allowed to learn and adjust them over time (and what you learn depends at least somewhat on what policies you choose). It turns out that learning doesn't help too much here, mostly because of a combination of slow learning rates and non-negligible economic discounting, unless you impose hard constraints (like "keep the probability of exceeding some temperature under x%", regardless of the economic costs). I'd like to try extending that work in several ways to see if that conclusion changes. One, I'd like to look at joint learning about several things: climate and carbon cycle feedbacks, climate and technological progress curves, etc. I'd also like to introduce more sophisticated notions of dynamics and learning (beyond simple Kalman filter updates of linear models with single scalar observations). I'm not sure if this is possible within existing dynamic programming frameworks, though.

    Beyond mitigation, I'd like to look at sequential decision making about climate adaptation, e.g. periodically deciding whether and how much to raise a sea wall. (I read a Masters thesis recently that looked at this in a simple way.) Or perhaps carbon stock management, e.g. trying to make sure your forest sequesters some minimum amount of carbon each year, which requires learning about your particular ecosystem's carbon uptake efficiency and sensitivity to climate change. In my current position I'm more likely to work on infrastructure adaptation problems. Ultimately I think my interests lie in areas other than climate mitigation, but it's hard to find problems where learning is relevant. (Either people make decisions frequently, every few years, so you don't need to anticipate what new information you'll receive; or they're made so rarely, 50+ years for some infrastructure, that you don't get much chance to update your beliefs.)

    I get the feeling everyone thinks I’m committed to analytically solvable models.

    I mostly work with non-analytic methods myself. The only reason I brought up analytic methods here is because some of them give simple and known solutions to somewhat non-trivial problems, so they might be a good place to get one's feet wet.

    Comment Source:John, > So, maybe it’s less important to detect ’tipping points’ than something like ’leverage points’: things to do that would have a big effect. This might be especially true if we could model mass behavior. Yes, although these are perhaps harder to identify and depend more strongly on socioeconomic assumptions. > What sort of project might you get into? I'm not sure yet. I have a few ideas of things to play around with. There has a bit of work on "endogenous learning" about climate feedbacks, i.e. how do mitigation policies change when you're allowed to learn and adjust them over time (and what you learn depends at least somewhat on what policies you choose). It turns out that learning doesn't help too much here, mostly because of a combination of slow learning rates and non-negligible economic discounting, unless you impose hard constraints (like "keep the probability of exceeding some temperature under x%", regardless of the economic costs). I'd like to try extending that work in several ways to see if that conclusion changes. One, I'd like to look at joint learning about several things: climate and carbon cycle feedbacks, climate and technological progress curves, etc. I'd also like to introduce more sophisticated notions of dynamics and learning (beyond simple Kalman filter updates of linear models with single scalar observations). I'm not sure if this is possible within existing dynamic programming frameworks, though. Beyond mitigation, I'd like to look at sequential decision making about climate adaptation, e.g. periodically deciding whether and how much to raise a sea wall. (I read a Masters thesis recently that looked at this in a simple way.) Or perhaps carbon stock management, e.g. trying to make sure your forest sequesters some minimum amount of carbon each year, which requires learning about your particular ecosystem's carbon uptake efficiency and sensitivity to climate change. In my current position I'm more likely to work on infrastructure adaptation problems. Ultimately I think my interests lie in areas other than climate mitigation, but it's hard to find problems where learning is relevant. (Either people make decisions frequently, every few years, so you don't need to anticipate what new information you'll receive; or they're made so rarely, 50+ years for some infrastructure, that you don't get much chance to update your beliefs.) > I get the feeling everyone thinks I’m committed to analytically solvable models. I mostly work with non-analytic methods myself. The only reason I brought up analytic methods here is because some of them give simple and known solutions to somewhat non-trivial problems, so they might be a good place to get one's feet wet.
Sign In or Register to comment.