Ongoing Projects

Opti­miza­tion prob­lems arise in many pro­duc­tion sectors, gen­er­ally with the aim of min­i­miz­ing costs and max­i­miz­ing profits. In Smart Cities and Indus­try 4.0, the effi­ciency of ser­vices pro­vided to cit­i­zens and effec­tive logis­tics for trans­port­ing people and goods become pri­or­i­ties, moti­vat­ing a range of aca­demic pro­pos­als to solve these prob­lems.

Facing the NP-hard nature of several com­bi­na­to­rial opti­miza­tion prob­lems rel­e­vant to these set­tings, such as the Data Mule Routing Problem in Mixed Net­works with Limited Auton­omy (PRM­DRM­cAL) and the Unca­pac­i­tated Network Design Problem with Short­est Path Con­straints (PCRNC-RCM), the project seeks both to advance the the­o­ret­i­cal under­stand­ing of the inter­ac­tion between Machine Learn­ing (ML) and Opti­miza­tion Methods (OM) and to offer prac­ti­cal solu­tions to press­ing chal­lenges in Smart Cities and Indus­try 4.0.

By inte­grat­ing ML with OM, the project is expected to open new pos­si­bil­i­ties for improv­ing the effi­ciency and effec­tive­ness of pro­duc­tion and logis­tics systems, con­tribut­ing to eco­nomic sus­tain­abil­ity and cit­i­zens’ quality of life. Suc­cess­ful hybrid solu­tions may change how highly complex opti­miza­tion prob­lems are tackled, with advances ranging from effi­cient resource man­age­ment to better urban ser­vices, in line with the goals of sus­tain­able devel­op­ment and tech­no­log­i­cal inno­va­tion.

The project carries out research and devel­op­ment (R&D) in part­ner­ship with the private sector to eval­u­ate, refor­mu­late and improve the math­e­mat­i­cal model cur­rently used to choose the matches broad­cast on free-to-air TV and on pay-per-view, taking into account audi­ence, regional balance, the sports cal­en­dar and com­mer­cial guide­lines.

The project carries out research and devel­op­ment (R&D) to con­sol­i­date and expand SID-Oceano as a national infra­struc­ture for oceano­graphic data and digital ser­vices, with empha­sis on the MVP of the Ocean Pre­dic­tion Centre (OCP), the MVP of the Exper­i­men­ta­tion Plat­form (PlatExp) and the Oceano­graphic Data Science Plat­form (PCDO), inte­grat­ing obser­va­tion, numer­i­cal mod­el­ling, data science and deliv­ery through dash­boards and APIs.

The theory of equi­lib­rium markets and its prac­ti­cal use in resource plan­ning has drawn growing inter­est because it models the inter­ac­tions among supply, demand and prices in inter­de­pen­dent systems. The project devel­ops com­pu­ta­tional tools to support the anal­y­sis, fore­cast­ing and opti­miza­tion of strate­gic deci­sions in complex markets for the Refin­ing, Trans­port and Trading (RTC) of oil and oil prod­ucts.

The work starts with a detailed under­stand­ing of the market equi­lib­rium problem, its the­o­ret­i­cal and prac­ti­cal fea­tures, and sim­pli­fied sce­nar­ios that serve as a basis for mod­el­ling. Math­e­mat­i­cal and com­pu­ta­tional models that capture the main dynam­ics of these markets are then built and imple­mented on a toy problem as a proof of concept.

After val­i­da­tion, the project moves on to real sce­nar­ios, includ­ing a rep­re­sen­ta­tive subset of crude oils and prod­ucts, explor­ing both exact and heuris­tic methods to solve the pro­posed models. Finally, the tools are adjusted to operate in a real RTC envi­ron­ment, con­sid­er­ing the par­tic­u­lar­i­ties of the sector, com­bin­ing math­e­mat­i­cal robust­ness with com­pu­ta­tional effi­ciency.

The project aims to (1) carry out the­o­ret­i­cal and prac­ti­cal research on the effec­tive inte­gra­tion of Machine Learn­ing tech­niques into Opti­miza­tion Methods; (2) develop hybrid methods for solving real instances of prob­lems arising in Smart Cities and Indus­try 4.0; (3) encour­age stu­dents to engage with the subject and give them the tools to work in this field; (4) foster col­lab­o­ra­tion among researchers from related areas within PESC/COPPE/UFRJ and partner insti­tu­tions, strength­en­ing exist­ing part­ner­ships and build­ing new ones; and (5) promote inter­na­tion­al­iza­tion, such as agree­ments and exchanges, through the con­tacts estab­lished during the research.

Opti­miza­tion prob­lems are common in indus­trial pro­duc­tion, focus­ing on min­i­miz­ing costs and max­i­miz­ing profits. In Smart Cities and Indus­try 4.0, service effi­ciency and trans­port logis­tics are pri­or­i­ties, moti­vat­ing aca­demic pro­pos­als to solve these prob­lems. Solving them exactly and quickly is essen­tial to the quality of pro­duc­tion chains, yet it remains a chal­lenge.

Com­bi­na­to­rial opti­miza­tion prob­lems have been widely studied because of their many prac­ti­cal appli­ca­tions. Being NP-hard, they call for meta-heuris­tics to find solu­tions quickly, but these can strug­gle on large prob­lems. The project there­fore devel­ops exact methods, meta-heuris­tics and hybrid methods that use Data Mining to handle real-world cases.

The main goal is to advance the state of the art in opti­miz­ing the con­fig­u­ra­tion of mon­i­tor­ing sensor net­works. A con­fig­u­ra­tion is eval­u­ated by the maximum entropy cri­te­rion, and its opti­miza­tion is mod­elled as the maximum entropy sam­pling problem (MESP), a highly chal­leng­ing mixed-integer non­lin­ear opti­miza­tion (MINLP) problem.

One line of work devel­ops an outer-approx­i­ma­tion algo­rithm for several integer non­lin­ear for­mu­la­tions of the MESP, com­bin­ing the best MINLP tech­niques with the best upper bounds for the MESP, based on convex relax­ations, through two main tools: dis­junc­tive pro­gram­ming and convex com­bi­na­tions of upper bounds.

Another line hybridizes scaling factors and masks, which can improve vir­tu­ally every upper-bound­ing method for the MESP. Opti­miza­tion over diag­o­nal masks is con­sid­ered, a gen­er­al­iza­tion of the scaling factor that remains a convex opti­miza­tion problem for some of the for­mu­la­tions studied.

Finally, the project studies exper­i­men­tal design from a broader per­spec­tive than the MESP, such as D-opti­mal­ity, which selects a subset of points of a linear regres­sion model to min­i­mize the gen­er­al­ized vari­ance of the least-squares esti­mates, and the gen­er­al­ized MESP, which extends both the MESP and D-opti­mal­ity.

The project devel­ops math­e­mat­i­cal models and exact and heuris­tic algo­rithms that inte­grate Opti­miza­tion and Machine Learn­ing to solve complex prob­lems arising from real appli­ca­tions in land trans­port logis­tics, off­shore logis­tics and health­care systems. It trains master’s and doc­toral stu­dents in the Trans­porta­tion Engi­neer­ing and Systems Engi­neer­ing and Com­puter Science grad­u­ate pro­grammes at COPPE/UFRJ and under­grad­u­ates at the UFRJ Poly­tech­nic School, gen­er­at­ing knowl­edge, tech­no­log­i­cal inno­va­tion and leading inter­na­tional sci­en­tific arti­cles.

The project devel­ops models and algo­rithms that inte­grate com­bi­na­to­rial opti­miza­tion, Markov deci­sion pro­cesses, machine learn­ing, sur­vival anal­y­sis and triage and primary care pro­to­cols, using his­tor­i­cal data on patient arrival rates, total oper­at­ing room capac­ity and the number of beds avail­able for post-sur­gi­cal recov­ery.

The goals are to map patient pro­files on entry into the hos­pi­tal system; iden­tify advanced-stage patients who should have been oper­ated on earlier; opti­mize surgery sched­ul­ing and bed man­age­ment con­sid­er­ing emer­gency flows; manage inter­nal flows to avoid can­cel­la­tions caused by system inef­fi­ciency and delayed pre-oper­a­tive exams; predict patient out­comes and length of stay; suggest pre-sur­gi­cal pro­to­cols for man­ag­ing comor­bidi­ties; suggest earlier refer­ral of emer­gency patients; and develop reha­bil­i­ta­tion pro­to­cols that con­sider patient char­ac­ter­is­tics and socioe­co­nomic context.

Expected results include the train­ing of new master’s and doc­toral grad­u­ates, high-impact pub­li­ca­tions, improve­ments to the real oper­a­tions studied and, con­se­quently, to patients’ phys­i­cal con­di­tion.

The project studies logis­tics prob­lems in trans­porta­tion engi­neer­ing that take into account envi­ron­men­tal and social aspects as well as tech­ni­cal and finan­cial ones. It devel­ops strong math­e­mat­i­cal models and exact and heuris­tic algo­rithms, with sen­si­tiv­ity anal­y­sis and Pareto opti­miza­tion, for com­bi­na­to­rial prob­lems of high com­pu­ta­tional com­plex­ity.

The real appli­ca­tions are: bus line fre­quency opti­miza­tion in light of over­crowd­ing dis­com­fort; capture of evasive flows on high­ways; loca­tion of traffic coun­ters; loca­tion of logis­tics inte­gra­tion centres con­sid­er­ing eco­nomic and envi­ron­men­tal aspects; loca­tion of tow trucks for highway inci­dents con­sid­er­ing con­ges­tion queues and envi­ron­men­tal aspects; and logis­tics dis­tri­bu­tion of hos­pi­tal sup­plies. Expected results include the train­ing of new master’s and doc­toral grad­u­ates, high-impact pub­li­ca­tions and improve­ments to the real oper­a­tions studied.

The main goal of this project is to explore novel ways of in-situ pro­cess­ing based on a smart storage plat­form devel­oped by the private sector. There are two main threads of work: (i) explor­ing SUCURI, a Python dataflow-based library, as an in-situ par­al­lel pro­gram­ming plat­form; and (ii) devel­op­ing in-situ arti­fi­cial intel­li­gence appli­ca­tions based on weight­less neural net­works.

The project devel­ops strong math­e­mat­i­cal models and exact and heuris­tic algo­rithms for com­bi­na­to­rial prob­lems of high com­pu­ta­tional com­plex­ity arising from real off­shore logis­tics appli­ca­tions: inven­tory con­sol­i­da­tion in the dynamic plan­ning of logis­tics net­works; peri­odic routing of cargo trans­port, het­ero­ge­neous fleet sizing, berth sched­ul­ing and the incor­po­ra­tion of meto­cean uncer­tainty into the plan­ning of off­shore support vessels; and the decom­mis­sion­ing of off­shore oil and gas facil­i­ties.

Expected results include the train­ing of new master’s and doc­toral grad­u­ates, high-impact pub­li­ca­tions and improve­ments to the real oper­a­tions studied.

The project devel­ops strong math­e­mat­i­cal models and exact and heuris­tic algo­rithms for com­bi­na­to­rial prob­lems of high com­pu­ta­tional com­plex­ity arising from real energy plan­ning appli­ca­tions: mod­el­ling the uncer­tainty of wind gen­er­a­tion in the daily oper­a­tion sched­ul­ing of the Brazil­ian National Inter­con­nected System (SIN); mod­el­ling exter­nal­i­ties and envi­ron­men­tal impacts in SIN gen­er­a­tion expan­sion plan­ning; and design­ing Demand Response mech­a­nisms for medium- and long-term expan­sion plan­ning.

Expected results include the train­ing of new master’s and doc­toral grad­u­ates, high-impact pub­li­ca­tions and improve­ments to the real oper­a­tions studied.

Special Vis­it­ing Researcher (PVE) project funded by CNPq, with the par­tic­i­pa­tion of Philippe Mich­e­lon.

This is the second phase of research into new ways of describ­ing and mod­el­ling real prob­lems as a com­po­si­tion of several prob­lems, pos­si­bly of dif­fer­ent kinds, in the form of pseudo-Boolean con­straints. Target prob­lems are tackled by syn­the­siz­ing an energy func­tion that rep­re­sents the state space of the problem’s solu­tions.

Several Com­pu­ta­tional Intel­li­gence models can be applied to search for global and, when needed, local minima of the func­tions pro­duced by the system. In par­tic­u­lar, the project seeks to take advan­tage of new quantum com­put­ers, specif­i­cally those per­form­ing adi­a­batic quantum com­pu­ta­tion. It also aims to gen­er­ate codes other than energy equa­tions, such as sharing graphs and weight­less neural net­works.

Project funded by CNPq, Uni­ver­sal Call, Tier C.

We study math­e­mat­i­cal opti­miza­tion and algo­rithm devel­op­ment for sta­tis­ti­cal models, focus­ing on sensor network appli­ca­tions. In these net­works, opti­miza­tion prob­lems arise nat­u­rally when select­ing subsets of sensors to acti­vate so as to max­i­mize the infor­ma­tion extracted from their mea­sure­ments. Infor­ma­tion is quan­ti­fied by dif­fer­en­tial entropy, which leads to the Maximum Entropy Sam­pling Problem (MESP).

The MESP is an NP-hard global opti­miza­tion problem with non-fac­torable objec­tive func­tions in non­lin­ear integer pro­gram­ming for­mu­la­tions. Despite this com­plex­ity, the par­tic­u­lar struc­ture of the objec­tive, the log-deter­mi­nant of a covari­ance matrix, has useful con­cav­ity and sub­mod­u­lar­ity prop­er­ties that enable effi­cient relax­ations and their use in exact branch-and-bound algo­rithms.

We also address opti­miza­tion methods for com­put­ing sparse gen­er­al­ized inverses, which are fun­da­men­tal to the effi­cient com­pu­ta­tion of many sta­tis­ti­cal quan­ti­ties and central to least-squares prob­lems. Both lines involve opti­miza­tion prob­lems of high rel­e­vance to sta­tis­tics and data science, and the project explores their con­cep­tual and algo­rith­mic con­nec­tions.

National Insti­tute of Science and Tech­nol­ogy sup­ported by CNPq and FAPERJ (CNPq grant 573768/2008-0).

Devel­op­ment of integer pro­gram­ming models (linear and non­lin­ear) and exact and heuris­tic algo­rithms for com­bi­na­to­rial prob­lems of high com­pu­ta­tional com­plex­ity arising from real appli­ca­tions in oil, energy plan­ning and health­care.