These are my recent Pin​board​.in links:

• Edge Per­spec­tives with John Hagel: Finite and Infi­nite Games — Which Game Shall We Play in the New Year?

  Far bet­ter, if pos­si­ble, to avoid direct con­fronta­tion and find ways to pur­sue infi­nite game play on the mar­gins or edges of finite game insti­tu­tions or in the white spaces not yet occu­pied by finite game insti­tu­tions.  By draw­ing atten­tion to hori­zons that have not yet been explored and demon­strat­ing the abil­ity to make progress in draw­ing out more poten­tial and pos­si­bil­ity, infi­nite game play­ers have a greater chance of shift­ing the game and attract­ing other play­ers. By build­ing par­al­lel insti­tu­tions and prac­tices that pull oth­ers into their game, infi­nite game play­ers can attract enough crit­i­cal mass so that they can pur­sue their quests with lower risk of inter­ven­tion from the finite game play­ers who view such actions as deeply sub­ver­sive.  At our research cen­ter, JSB and I are now explor­ing these kinds of approaches as a way of achiev­ing orga­ni­za­tional change within large institutions.
  
  what-​​I-​​do

• [1107.0056] Fixed para­me­ter algo­rithms for restricted col­or­ing problems

  In this paper, we obtain poly­no­mial time algo­rithms to deter­mine the acyclic chro­matic num­ber, the star chro­matic num­ber, the Thue chro­matic num­ber, the har­mo­nious chro­matic num­ber and the clique chro­matic num­ber of $P_4$-tidy graphs and $(q,q-4)$-graphs, for every fixed $q$. These classes include cographs, $P_4$-sparse and $P_4$-lite graphs. All these col­or­ing prob­lems are known to be NP-​​hard for gen­eral graphs. These algo­rithms are fixed para­me­ter tractable on the para­me­ter $q(G)$, which is the min­i­mum $q$ such that $G$ is a $(q,q-4)$-graph. We also prove that every con­nected $(q,q-4)$-graph with at least $q$ ver­tices is 2-​​clique-​​colorable and that every acyclic col­or­ing of a cograph is also nonrepetitive.
  
  algo­rithms graph-​​theory discrete-​​mathematics nudge-​​targets

• [1112.6045] Com­par­ing inter­mit­tency and net­work mea­sure­ments of words and their depen­dency on authorship

  Many fea­tures from texts and lan­guages can now be inferred from sta­tis­ti­cal analy­ses using con­cepts from com­plex net­works and dynam­i­cal sys­tems. In this paper we quan­tify how topo­log­i­cal prop­er­ties of word co-​​occurrence net­works and inter­mit­tency (or bursti­ness) in word dis­tri­b­u­tion depend on the style of authors. Our data­base con­tains 40 books from 8 authors who lived in the 19^th and 20^th cen­turies, for which the fol­low­ing net­work mea­sure­ments were obtained: clus­ter­ing coef­fi­cient, aver­age short­est path lengths, and between­ness. We found that the two fac­tors with stronger depen­dency on the authors were the skew­ness in the dis­tri­b­u­tion of word inter­mit­tency and the aver­age short­est paths. Other fac­tors such as the betwee­ness and the Zipf’s law expo­nent show only weak depen­dency on author­ship. Also assessed was the con­tri­bu­tion from each mea­sure­ment to author­ship recog­ni­tion using three machine learn­ing meth­ods. The best per­for­mance was a ca. 65 % accu­racy upon com­bin­ing com­plex net­work and inter­mit­tency fea­tures with the near­est neigh­bor algo­rithm. From a detailed analy­sis of the inter­de­pen­dence of the var­i­ous met­rics it is con­cluded that the meth­ods used here are com­ple­men­tary for pro­vid­ing short– and long-​​scale per­spec­tives of texts, which are use­ful for appli­ca­tions such as iden­ti­fi­ca­tion of top­i­cal words and infor­ma­tion retrieval.
  
  natural-​​language-​​processing document-​​clustering clus­ter­ing feature-​​selection algo­rithms nudge-​​targets

• [1108.1170] Con­vex Opti­miza­tion with­out Pro­jec­tion Steps

  For the gen­eral prob­lem of min­i­miz­ing a con­vex func­tion over a com­pact con­vex domain, we will inves­ti­gate a sim­ple iter­a­tive approx­i­ma­tion algo­rithm based on the method by Frank & Wolfe 1956, that does not need pro­jec­tion steps in order to stay inside the opti­miza­tion domain. Instead of a pro­jec­tion step, the lin­earized prob­lem defined by a cur­rent sub­gra­di­ent is solved, which gives a step direc­tion that will nat­u­rally stay in the domain. Our frame­work gen­er­al­izes the sparse greedy algo­rithm of Frank & Wolfe and its primal-​​dual analy­sis by Clark­son 2010 (and the low-​​rank SDP approach by Hazan 2008) to arbi­trary con­vex domains. We give a con­ver­gence proof guar­an­tee­ing {epsilon}-small dual­ity gap after O(1/{epsilon}) iter­a­tions. The method allows us to under­stand the spar­sity of approx­i­mate solu­tions for any l1-​​regularized con­vex opti­miza­tion prob­lem (and for opti­miza­tion over the sim­plex), expressed as a func­tion of the approx­i­ma­tion qual­ity. We obtain match­ing upper and lower bounds of {Theta}(1/{epsilon}) for the spar­sity for l1-​​problems. The same bounds apply to low-​​rank semi­def­i­nite opti­miza­tion with bounded trace, show­ing that rank O(1/{epsilon}) is best pos­si­ble here as well. As another appli­ca­tion, we obtain sparse matri­ces of O(1/{epsilon}) non-​​zero entries as {epsilon}-approximate solu­tions when opti­miz­ing any con­vex func­tion over a class of diag­o­nally dom­i­nant sym­met­ric matri­ces. We show that our pro­posed first-​​order method also applies to nuclear norm and max-​​norm matrix opti­miza­tion prob­lems. For nuclear norm reg­u­lar­ized opti­miza­tion, such as matrix com­ple­tion and low-​​rank recov­ery, we demon­strate the prac­ti­cal effi­ciency and scal­a­bil­ity of our algo­rithm for large matrix prob­lems, as e.g. the Net­flix dataset. For gen­eral con­vex opti­miza­tion over bounded matrix max-​​norm, our algo­rithm is the first with a con­ver­gence guar­an­tee, to the best of our knowledge.
  
  operations-​​research opti­miza­tion algo­rithms nudge-​​targets

• [1112.6235] Detect­ing a Vec­tor Based on Lin­ear Measurements

  We con­sider a sit­u­a­tion where the state of a sys­tem is rep­re­sented by a real-​​valued vec­tor. Under nor­mal cir­cum­stances, the vec­tor is zero, while an event man­i­fests as non-​​zero entries in this vec­tor, pos­si­bly few. Our inter­est is in the design of algo­rithms that can reli­ably detect events (i.e., test whether the vec­tor is zero or not) with the least amount of infor­ma­tion. We place our­selves in a sit­u­a­tion, now com­mon in the sig­nal pro­cess­ing lit­er­a­ture, where infor­ma­tion about the vec­tor comes in the form of noisy lin­ear mea­sure­ments. We derive infor­ma­tion bounds in an active learn­ing setup and exhibit some sim­ple near-​​optimal algo­rithms. In par­tic­u­lar, our results show that the task of detec­tion within this set­ting is at once much eas­ier, sim­pler and dif­fer­ent than the tasks of esti­ma­tion and sup­port recovery.
  
  signal-​​processing sta­tis­tics algo­rithms nudge-​​targets

• [1109.2215] Find­ing miss­ing edges and com­mu­ni­ties in incom­plete networks

  Many algo­rithms have been pro­posed for pre­dict­ing miss­ing edges in net­works, but they do not usu­ally take account of which edges are miss­ing. We focus on net­works which have miss­ing edges of the form that is likely to occur in real net­works, and com­pare algo­rithms that find these miss­ing edges. We also inves­ti­gate the effect of this kind of miss­ing data on com­mu­nity detec­tion algorithms.
  
  network-​​theory algo­rithms infer­ence sta­tis­tics nudge-​​targets

• [1010.4735] Explor­ing the Energy Land­scapes of Pro­tein Fold­ing Sim­u­la­tions with Bayesian Computation

  Nested sam­pling is a Bayesian sam­pling tech­nique devel­oped to explore prob­a­bil­ity dis­tri­b­u­tions lo– calised in an expo­nen­tially small area of the para­me­ter space. The algo­rithm pro­vides both pos­te­rior sam­ples and an esti­mate of the evi­dence (mar­ginal like­li­hood) of the model. The nested sam­pling algo– rithm also pro­vides an effi­cient way to cal­cu­late free ener­gies and the expec­ta­tion value of ther­mo­dy­namic observ­ables at any tem­per­a­ture, through a sim­ple post-​​processing of the out­put. Pre­vi­ous appli­ca­tions of the algo­rithm have yielded large effi­ciency gains over other sam­pling tech­niques, includ­ing par­al­lel tem­per­ing (replica exchange). In this paper we describe a par­al­lel imple­men­ta­tion of the nested sam­pling algo­rithm and its appli­ca­tion to the prob­lem of pro­tein fold­ing in a Go-​​type force field of empir­i­cal poten­tials that were designed to sta­bi­lize sec­ondary struc­ture ele­ments in room-​​temperature sim­u­la­tions. We demon­strate the method by con­duct­ing fold­ing sim­u­la­tions on a num­ber of small pro­teins which are com­monly used for test­ing pro­tein fold­ing pro­ce­dures: pro­tein G, the SH3 domain of Src tyro­sine kinase and chy­motrypsin inhibitor 2. A topo­log­i­cal analy­sis of the pos­te­rior sam­ples is per­formed to pro­duce energy land­scape charts, which give a high level descrip­tion of the poten­tial energy sur­face for the pro­tein fold­ing sim­u­la­tions. These charts pro­vide qual­i­ta­tive insights into both the fold­ing process and the nature of the model and force field used.
  
  structural-​​biology bio­chem­istry mod­el­ing algo­rithms sta­tis­tics meta­mod­el­ing

• [1109.2618] Fast and Accu­rate Mod­el­ing of Mol­e­c­u­lar Atom­iza­tion Ener­gies with Machine Learning

  We intro­duce a machine learn­ing model to pre­dict atom­iza­tion ener­gies of a diverse set of organic mol­e­cules, based on nuclear charges and atomic posi­tions only. The prob­lem of solv­ing the mol­e­c­u­lar Schr“odinger equa­tion is mapped onto a non-​​linear sta­tis­ti­cal regres­sion prob­lem of reduced com­plex­ity. Regres­sion mod­els are trained on and com­pared to atom­iza­tion ener­gies com­puted with hybrid density-​​functional the­ory. Cross-​​validation over more than seven thou­sand small organic mol­e­cules yields a mean absolute error of ~10 kcal/​mol. Applic­a­bil­ity is demon­strated for the pre­dic­tion of mol­e­c­u­lar atom­iza­tion poten­tial energy curves.
  
  machine-​​learning learning-​​from-​​data bio­chem­istry computational-​​science nudge-​​targets

• [1101.2135] Bounded con­fi­dence model: addressed infor­ma­tion main­tain diver­sity of opinions

  A com­mu­nity of agents is sub­ject to a stream of mes­sages, which are rep­re­sented as points on a plane of issues. Mes­sages are sent by media and by agents them­selves. Mes­sages from media shape the pub­lic opin­ion. They are unbi­ased, i.e. pos­i­tive and neg­a­tive opin­ions on a given issue appear with equal fre­quen­cies. In our pre­vi­ous work, the only cri­te­rion to receive a mes­sage by an agent is if the dis­tance between this mes­sage and the ones received ear­lier does not exceed the given value of the tol­er­ance para­me­ter. Here we intro­duce a pos­si­bil­ity to address a mes­sage to a given neigh­bour. We show that this option reduces the una­nim­ity effect, what improves the col­lec­tive performance.
  
  agent-​​based com­mu­ni­ca­tion network-​​theory machine-​​learning diver­sity