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

• [1106.1804] A Crit­i­cal Assess­ment of Bench­mark Com­par­i­son in Planning

  “Recent trends in plan­ning research have led to empir­i­cal com­par­i­son becom­ing com­mon­place. The field has started to set­tle into a method­ol­ogy for such com­par­isons, which for obvi­ous prac­ti­cal rea­sons requires run­ning a sub­set of plan­ners on a sub­set of prob­lems. In this paper, we char­ac­ter­ize the method­ol­ogy and exam­ine eight implicit assump­tions about the prob­lems, plan­ners and met­rics used in many of these com­par­isons. The prob­lem assump­tions are: PR1) the per­for­mance of a gen­eral pur­pose plan­ner should not be penalized/​biased if exe­cuted on a sam­pling of prob­lems and domains, PR2) minor syn­tac­tic dif­fer­ences in rep­re­sen­ta­tion do not affect per­for­mance, and PR3) prob­lems should be solv­able by STRIPS capa­ble plan­ners unless they require ADL. The plan­ner assump­tions are: PL1) the lat­est ver­sion of a plan­ner is the best one to use, PL2) default para­me­ter set­tings approx­i­mate good per­for­mance, and PL3) time cut-​​offs do not unduly bias out­come. The met­rics assump­tions are: M1) per­for­mance degrades sim­i­larly for each plan­ner when run on degraded run­time envi­ron­ments (e.g., machine plat­form) and M2) the num­ber of plan steps dis­tin­guishes per­for­mance. We find that most of these assump­tions are not sup­ported empir­i­cally; in par­tic­u­lar, that plan­ners are affected dif­fer­ently by these assump­tions. We con­clude with a call to the com­mu­nity to devote research resources to improv­ing the state of the prac­tice and espe­cially to enhanc­ing the avail­able bench­mark problems.”
  
  plan­ning bench­mark­ing algo­rithms horse-​​races engineering-​​design operations-​​research nudge-​​targets

• [1108.4361] The rela­tion­ship between acquain­tance­ship and coau­thor­ship in sci­en­tific col­lab­o­ra­tion networks

  “This arti­cle exam­ines the rela­tion­ship between acquain­tance­ship and coau­thor­ship pat­terns in a multi-​​disciplinary, multi-​​institutional, geo­graph­i­cally dis­trib­uted research cen­ter. Two social net­works are con­structed and com­pared: a net­work of coau­thor­ship, rep­re­sent­ing how researchers write arti­cles with one another, and a net­work of acquain­tance­ship, rep­re­sent­ing how those researchers know each other on a per­sonal level, based on their responses to an online sur­vey. Sta­tis­ti­cal analy­ses of the topol­ogy and com­mu­nity struc­ture of these net­works point to the impor­tance of small-​​scale, local, per­sonal net­works pred­i­cated upon acquain­tance­ship for accom­plish­ing col­lab­o­ra­tive work in sci­en­tific communities.”
  
  academic-​​culture network-​​theory cita­tion social-​​networks

• [1108.4223] The set-​​theoretic multiverse

  “The mul­ti­verse view in set the­ory, intro­duced and argued for in this arti­cle, is the view that there are many dis­tinct con­cepts of set, each instan­ti­ated in a cor­re­spond­ing set-​​theoretic uni­verse. The uni­verse view, in con­trast, asserts that there is an absolute back­ground set con­cept, with a cor­re­spond­ing absolute set-​​theoretic uni­verse in which every set-​​theoretic ques­tion has a def­i­nite answer. The mul­ti­verse posi­tion, I argue, explains our expe­ri­ence with the enor­mous diver­sity of set-​​theoretic pos­si­bil­i­ties, a phe­nom­e­non that chal­lenges the uni­verse view. In par­tic­u­lar, I argue that the con­tin­uum hypoth­e­sis is set­tled on the mul­ti­verse view by our exten­sive knowl­edge about how it behaves in the mul­ti­verse, and as a result it can no longer be set­tled in the man­ner for­merly hoped for.”
  
  math­e­mat­ics mathematical-​​criticism looking-​​forward-​​to-​​understanding-​​this-​​someday pragmatism-it-ain’t

• [1102.1934] The struc­ture of the Arts & Human­i­ties Cita­tion Index: A map­ping on the basis of aggre­gated cita­tions among 1,157 journals

  “Using the Arts & Human­i­ties Cita­tion Index (A&HCI) 2008, we apply map­ping tech­niques pre­vi­ously devel­oped for map­ping jour­nal struc­tures in the Sci­ence and Social Sci­ence Cita­tion Indices. Cita­tion rela­tions among the 110,718 records were aggre­gated at the level of 1,157 jour­nals spe­cific to the A&HCI, and the jour­nal struc­tures are ques­tioned on whether a cog­ni­tive struc­ture can be recon­structed and visu­al­ized. Both cosine-​​normalization (bot­tom up) and fac­tor analy­sis (top down) sug­gest a divi­sion into approx­i­mately twelve sub­sets. The rela­tions among these sub­sets are explored using var­i­ous visu­al­iza­tion tech­niques. How­ever, we were not able to retrieve this struc­ture using the ISI Sub­ject Cat­e­gories, includ­ing the 25 cat­e­gories which are spe­cific to the A&HCI. We dis­cuss options for val­i­da­tion such as against the cat­e­gories of the Human­i­ties Indi­ca­tors of the Amer­i­can Acad­emy of Arts and Sci­ences, the panel struc­ture of the Euro­pean Ref­er­ence Index for the Human­i­ties (ERIH), and com­pare our results with the cur­ricu­lum orga­ni­za­tion of the Human­i­ties Sec­tion of the Col­lege of Let­ters and Sci­ences of UCLA as an exam­ple of insti­tu­tional organization.”
  
  network-​​theory citation-​​networks human­i­ties academic-​​culture quantitative-​​humanities

• [1108.4220] A Dynam­i­cal Sys­tems Approach for Sta­tic Eval­u­a­tion in Go

  “In the paper argu­ments are given why the con­cept of sta­tic eval­u­a­tion has the poten­tial to be a use­ful exten­sion to Monte Carlo tree search. A new con­cept of mod­el­ing sta­tic eval­u­a­tion through a dynam­i­cal sys­tem is intro­duced and strengths and weak­nesses are dis­cussed. The gen­eral suit­abil­ity of this approach is demonstrated.”
  
  representation-​​theory plan­ning monte-​​carlo-​​models nudge algo­rithms

• [1105.5449] AntNet: Dis­trib­uted Stig­mer­getic Con­trol for Com­mu­ni­ca­tions Networks

  “…We com­pare our algo­rithm with six state-​​of-​​the-​​art rout­ing algo­rithms com­ing from the telecom­mu­ni­ca­tions and machine learn­ing fields. The algo­rithms’ per­for­mance is eval­u­ated over a set of real­is­tic test­beds. We run many exper­i­ments over real and arti­fi­cial IP data­gram net­works with increas­ing num­ber of nodes and under sev­eral par­a­dig­matic spa­tial and tem­po­ral traf­fic dis­tri­b­u­tions. Results are very encour­ag­ing. AntNet showed supe­rior per­for­mance under all the exper­i­men­tal con­di­tions with respect to its com­peti­tors. We ana­lyze the main char­ac­ter­is­tics of the algo­rithm and try to explain the rea­sons for its superiority.”
  
  ant-​​colony-​​optimization network-​​theory net­works con­trol algo­rithms nudge-​​targets rout­ing

• Bozo Sapi­ens: Sacco and Vanzetti: Evidence

  “Wigmore’s tech­nique, like prob­a­bil­ity itself, is both wide-​​ranging and tediously painstak­ing; his book was pop­u­lar only among insom­niac judges. But now that com­put­ers can take on the numer­i­cal drudgery, it is prov­ing its worth in just such tan­gled cases as Sacco’s and Vanzetti’s. The legal schol­ars Joseph Kadane and David Schum have applied a sophis­ti­cated exten­sion of Wigmore’s method to the vast body of evi­dence from the case. Theirs is a remark­able achieve­ment; their charts retain all the orig­i­nal com­plex­i­ties: the facts with­held or per­verted, the hearsay, the lies told and dis­avowed on both sides, the charged polit­i­cal atmos­phere of eighty years ago. They never dis­count a fact, no mat­ter how far-​​fetched; they  sim­ply give it its due weight in their dynamic struc­ture. Their con­clu­sion?  Unjust though it is to sum­ma­rize a book in a sen­tence, the bal­ance of prob­a­bil­ity seems to favor the view expressed long ago by one of the defen­dants’ close com­pan­ions: “every­one in the Boston anar­chis­tic cir­cle knew that Sacco was guilty and that Vanzetti was inno­cent as far as the actual par­tic­i­pa­tion in the killing.” So, there it is: whichever side our polit­i­cal instincts favor, we are des­tined to be half wrong. Vanzetti’s last words were: “I wish to for­give some peo­ple for what they are now doing to me.”  If we were all will­ing to make the extra effort to work out the prob­a­bil­i­ties, per­haps we might not need for­give­ness so often.”
  
  probability-​​theory legal-​​studies computational-​​methods his­tory

• Get­ting first sale wrong

  “I hate to imag­ine it, but this deci­sion raises some fright­en­ing pos­si­bil­i­ties and requires greater vig­i­lance on the part of librar­i­ans.  At the very least, libraries must demand infor­ma­tion from pub­lish­ers about where every item has been man­u­fac­tured. Obtain­ing such infor­ma­tion is no longer an option, since our legal uses of the things we buy now depends on know­ing this, and the place where the pub­lisher is located or where the sale took place is sim­ply not suf­fi­cient.  But what I really fear is that pub­lish­ers will begin to man­u­fac­ture more of their works over­seas and then try to demand a higher price – one that includes “pub­lic lend­ing rights” – from libraries. If libraries are in a dif­fi­cult posi­tion, stu­dents may be even worse off under the Sec­ond Circuit’s rul­ing.  Again, pub­lish­ers now have an incen­tive to man­u­fac­ture their text­books abroad and sell them to U.S. stu­dents.  Such stu­dents would no longer have the right to re-​​sell their text­books or to pur­chase used texts.  The defen­dant in the case, Supap Kirt­saeng, had made a lucra­tive busi­ness out of reselling text­books pur­chased in Asia.  He was per­haps an unsym­pa­thetic party, but what he was doing was not dif­fer­ent in kind from the resale of texts that is com­mon on all col­lege cam­puses.  This activ­ity makes higher edu­ca­tion a lit­tle more pos­si­ble for many.  Now pub­lish­ers have an easy way for to close down this sec­ondary mar­ket for text­books, about which they have com­plained for years.  In the process, the cost of edu­ca­tion for col­lege stu­dents would be pushed up even further.”
  
  copy­right insan­ity intellectual-​​property academic-​​culture librar­i­ans

• [1106.6037] Black Hole Search with Finite Automata Scat­tered in a Syn­chro­nous Torus

  “We con­sider the prob­lem of locat­ing a black hole in syn­chro­nous anony­mous net­works using finite state agents. A black hole is a harm­ful node in the net­work that destroys any agent vis­it­ing that node with­out leav­ing any trace. The objec­tive is to locate the black hole with­out destroy­ing too many agents. This is dif­fi­cult to achieve when the agents are ini­tially scat­tered in the net­work and are unaware of the loca­tion of each other. Pre­vi­ous stud­ies for black hole search used more pow­er­ful mod­els where the agents had non-​​constant mem­ory, were labelled with dis­tinct iden­ti­fiers and could either write mes­sages on the nodes of the net­work or mark the edges of the net­work. In con­trast, we solve the prob­lem using a small team of finite-​​state agents each car­ry­ing a con­stant num­ber of iden­ti­cal tokens that could be placed on the nodes of the net­work. Thus, all resources used in our algo­rithms are inde­pen­dent of the net­work size. We restrict our atten­tion to ori­ented torus net­works and first show that no finite team of finite state agents can solve the prob­lem in such net­works, when the tokens are not mov­able. In case the agents are equipped with mov­able tokens, we deter­mine lower bounds on the num­ber of agents and tokens required for solv­ing the prob­lem in torus net­works of arbi­trary size. Fur­ther, we present a deter­min­is­tic solu­tion to the black hole search prob­lem for ori­ented torus net­works, using the min­i­mum num­ber of agents and tokens.”
  
  algo­rithms agent-​​based multi-​​agent-​​systems network-​​theory nudge-​​targets

• [1106.1821] Col­lec­tive Intel­li­gence, Data Rout­ing and Braess’ Paradox

  “We con­sider the prob­lem of design­ing the the util­ity func­tions of the utility-​​maximizing agents in a multi-​​agent sys­tem so that they work syn­er­gis­ti­cally to max­i­mize a global util­ity. The par­tic­u­lar prob­lem domain we explore is the con­trol of net­work rout­ing by plac­ing agents on all the routers in the net­work. Con­ven­tional approaches to this task have the agents all use the Ideal Short­est Path rout­ing Algo­rithm (ISPA). We demon­strate that in many cases, due to the side-​​effects of one agent’s actions on another agent’s per­for­mance, hav­ing agents use ISPA’s is sub­op­ti­mal as far as global aggre­gate cost is con­cerned, even when they are only used to route infin­i­tes­i­mally small amounts of traf­fic. The util­ity func­tions of the indi­vid­ual agents are not “aligned” with the global util­ity, intu­itively speak­ing. As a par­tic­u­lar exam­ple of this we present an instance of Braess’ para­dox in which adding new links to a net­work whose agents all use the ISPA results in a decrease in over­all through­put. We also demon­strate that load-​​balancing, in which the agents’ deci­sions are col­lec­tively made to opti­mize the global cost incurred by all traf­fic cur­rently being routed, is sub­op­ti­mal as far as global cost aver­aged across time is con­cerned. This is also due to ‘side-​​effects’, in this case of cur­rent rout­ing deci­sion on future traf­fic. The math­e­mat­ics of Col­lec­tive Intel­li­gence (COIN) is con­cerned pre­cisely with the issue of avoid­ing such dele­te­ri­ous side-​​effects in multi-​​agent sys­tems, both over time and space. We present key con­cepts from that math­e­mat­ics and use them to derive an algo­rithm whose ideal ver­sion should have bet­ter per­for­mance than that of hav­ing all agents use the ISPA, even in the infin­i­tes­i­mal limit. We present exper­i­ments ver­i­fy­ing this, and also show­ing that a machine-​​learning-​​based ver­sion of this COIN algo­rithm in which costs are only impre­cisely esti­mated via empir­i­cal means (a ver­sion poten­tially applic­a­ble in the real world) also out­per­forms the ISPA, despite hav­ing access to less infor­ma­tion than does the ISPA. In par­tic­u­lar, this COIN algo­rithm almost always avoids Braess’ paradox.”
  
  collective-​​intelligence search-​​algorithms figure-​​ground-​​error plan­ning nudge

• [1108.0404] Exploit­ing Agent and Type Inde­pen­dence in Col­lab­o­ra­tive Graph­i­cal Bayesian Games

  “Effi­cient col­lab­o­ra­tive deci­sion mak­ing is an impor­tant chal­lenge for mul­ti­a­gent sys­tems. Find­ing opti­mal joint actions is espe­cially chal­leng­ing when each agent has only imper­fect infor­ma­tion about the state of its envi­ron­ment. Such prob­lems can be mod­eled as col­lab­o­ra­tive Bayesian games in which each agent receives pri­vate infor­ma­tion in the form of its type. How­ever, rep­re­sent­ing and solv­ing such games requires space and com­pu­ta­tion time expo­nen­tial in the num­ber of agents. This arti­cle intro­duces col­lab­o­ra­tive graph­i­cal Bayesian games (CGBGs), which facil­i­tate more effi­cient col­lab­o­ra­tive deci­sion mak­ing by decom­pos­ing the global pay­off func­tion as the sum of local pay­off func­tions that depend on only a few agents. We pro­pose a frame­work for the effi­cient solu­tion of CGBGs based on the insight that they posses two dif­fer­ent types of inde­pen­dence, which we call agent inde­pen­dence and type inde­pen­dence. In par­tic­u­lar, we present a fac­tor graph rep­re­sen­ta­tion that cap­tures both forms of inde­pen­dence and thus enables effi­cient solu­tions. In addi­tion, we show how this rep­re­sen­ta­tion can pro­vide lever­age in sequen­tial tasks by using it to con­struct a novel method for decen­tral­ized par­tially observ­able Markov deci­sion processes. Exper­i­men­tal results in both ran­dom and bench­mark tasks demon­strate the improved scal­a­bil­ity of our meth­ods com­pared to sev­eral exist­ing alternatives.”
  
  col­lab­o­ra­tion agent-​​based complex-​​systems emergent-​​design nudge-​​targets

• [1102.2837] Effi­cient Pro­mo­tion Strate­gies in Hier­ar­chi­cal Organizations

  “The Peter prin­ci­ple has been recently inves­ti­gated by means of an agent-​​based sim­u­la­tion and its valid­ity has been numer­i­cally cor­rob­o­rated. It has been con­firmed that, within cer­tain con­di­tions, it can really influ­ence in a neg­a­tive way the effi­ciency of a pyra­mi­dal orga­ni­za­tion adopt­ing mer­i­to­cratic pro­mo­tions. It was also found that, in order to bypass these effects, alter­na­tive pro­mo­tion strate­gies should be adopted, as for exam­ple a ran­dom selec­tion choice. In this paper, within the same line of research, we study pro­mo­tion strate­gies in a more real­is­tic hier­ar­chi­cal and mod­u­lar orga­ni­za­tion and we show the robust­ness of our pre­vi­ous results, extend­ing their valid­ity to a more gen­eral con­text. We dis­cuss also why the adop­tion of these strate­gies could be use­ful for real organizations.”
  
  organizational-​​behavior com­plex­ol­ogy complexological-​​amusements agent-​​based com­pe­tence