Roadmap

[relf]

Some ideas for further developments:

• Make infill criterion a trait to allow to use EgorSolver with custom criterion (Make the infill criterion a trait #92 )
• Implement gaussian process sampling (Gpx sampling #97 )
• Put nlopt behind feature gate to get pure rust implementation (GP uses pure Rust COBYLA #110, EGO and then egobox as pure Rust implementation #112 )
• Implement ask-and-tell interface (Egor ask-and-tell interface #116 )
• Manage noisy training data (sparse GP Implement sparse gaussian process methods #128, Rework SGP API #129, Rework Python GpMix API #130, Mixture of sparse GP #131)
• Make some benchmarks vs SMT (https://github.com/Antoine-Averland/benchmark_smt_egobox)
• Implement TREGO algorithm (Implement TREGO algorithm #173)
• Implement hotstart using checkpointing argmin mechanism (Add checkpointing and hot start #197)
• Implement criterion taking into account constraints (UTB Implement upper trusted bound criterion for constraint handling #245 cf. paper , cEI Implement constrained infill criterion #246 cf. paper)
• Implement CoEGO (cf. paper) (CoEGO implementation and refactoring #253, Make CoEGO activity matrix more balanced #265)
• Experiment GBNM infill criterion optimizer (cf. paper) (Add experimental Global Bounded Nelder-Mead infill criterion optimizer  #264) finally removed in Remove experimental GBNM internal optimizer #285
• Add LogEI infill criterion (cf. paper) (Implement LogCEI criterion #284, Use Logarithm of Expected Improvement (LogEi) as the default infill strategy #336)
• Handle hidden (aka unknown, crash) constraints (cf. paper) (Implementation of hidden constraints handling #377, Add probability of viability constraint as failsafe strategy #378)
• Make mixed integer surrogate available (GpMix accepts now xspecs to create Gpx from mixed-integer variables #388 )
• Handle other constraint format (not only <=0) (Handle other constraint format (not only c <= 0) #414)
• [ ] Implement power_1.9_exponential kernel to avoid ill-conditionning with small theta values cf. paper. Finally ill-conditionning occurs for very small theta outside default theta bounds, so not worth implementing for now.
• Implement Feasibility Enhanced criteria cf. paper
• ...