LocalSolver 4.0 Beta : discover new features
Forum 'Annonces' - Sujet créé le 2013-12-07 par Frédéric Gardi
For people interested in large-scale mixed-variable non-convex optimization, the Beta version of LocalSolver 4.0 is available! Please have a look to http://www.localsolver.com/betaversions.html The Beta 4.0 version is only provided for Windows platforms. The release version provided for all platforms will be available before Christmas.
The main new features are:
- Continuous decisions can now be declared in addition to binary decisions. Just use the operator "float(lower, upper)" instead of "bool()", by specifying the lower and upper bounds of the continuous decision. This feature allows to tackle large-scale continuous or mixed-variable non-convex optimization problems.
For fast and scalable optimization over continuous domains, we have integrated small-neighborhood exploration techniques dedicated to continuous domains. These techniques are related to the ones used to optimize over combinatorial spaces in LocalSolver 3.1. For continuous domains, these techniques are related to derivative-free and direct search techniques as used in non-linear, non-convex, non-smooth, or non-differentiable optimization.
Next LocalSolver versions will be enriched by large-scale neighborhood explorations related to sucessive linear or quadratic programming techniques (SLP, SQP) well-known in the non-linear programming world.
- The preprocessing phase has been improved to tackle models which are not suited for LocalSolver, in particular models originally modeled for mixed-integer linear programming solvers. This preprocessing phase provides lower bounds for each objective through constraint propagation and inference techniques. These lower bounds are provided at the beginning of the trace when using the LSP language. Having improved these inference techniques, LocalSolver 4.0 is able to prove optimality or infeasibility on more problems than with 3.1 version.
- To search and optimize over combinatorial domains, large-neighborhood explorations based on mixed-integer linear programming techniques are now available to diversify the search and to escape local optima related to small neighborhoods, to ultimately find better solutions. It consists essentially in exploiting the linear relaxation of appropriate subproblems, coupled with primal rounding heuristics to find quickly good-quality integer solutions. In this way, when neighborhoods (that is, subproblems) have a good linear relaxation, LocalSolver is now able to exploit it. The linear relaxation is computed using a revised dual simplex algorithm.
Minor features:
- LocalSolver now automatically manages impossible values for some operators: zero value for division, negative values for square root, zero value for logarithm, etc. With such impossible values in input, the operator outputs Not-A-Number (NaN) as value. LocalSolver automatically searches for solutions not leading to NaN values for these operators.
- LocalSolver now automatically manages an empty list of operands for n-ary operators like: sum, prod, or, and, xor. The value returned by an empty expression corresponds to the neutral value for the corresponding operator. For example: for sum the neutral value is 0; for prod it is 1; for or it is false or 0; for and it is true or 1; for xor it is false or 0.
Among the classes of problems you can newly tackle with LocalSolver are: supply chain optimization problems, unit commitment problems, portfolio optimization problems, as well as diverse numerical optimization problems. The main advantage of using LocalSolver over other mathematical optimization solvers of the market is to make you able to tackle ultra-large problems involving millions of variables within short computation times.
LocalSolver is free for academics: create your account in one minute on http://www.localsolver.com/account.html to try it!
The main new features are:
- Continuous decisions can now be declared in addition to binary decisions. Just use the operator "float(lower, upper)" instead of "bool()", by specifying the lower and upper bounds of the continuous decision. This feature allows to tackle large-scale continuous or mixed-variable non-convex optimization problems.
For fast and scalable optimization over continuous domains, we have integrated small-neighborhood exploration techniques dedicated to continuous domains. These techniques are related to the ones used to optimize over combinatorial spaces in LocalSolver 3.1. For continuous domains, these techniques are related to derivative-free and direct search techniques as used in non-linear, non-convex, non-smooth, or non-differentiable optimization.
Next LocalSolver versions will be enriched by large-scale neighborhood explorations related to sucessive linear or quadratic programming techniques (SLP, SQP) well-known in the non-linear programming world.
- The preprocessing phase has been improved to tackle models which are not suited for LocalSolver, in particular models originally modeled for mixed-integer linear programming solvers. This preprocessing phase provides lower bounds for each objective through constraint propagation and inference techniques. These lower bounds are provided at the beginning of the trace when using the LSP language. Having improved these inference techniques, LocalSolver 4.0 is able to prove optimality or infeasibility on more problems than with 3.1 version.
- To search and optimize over combinatorial domains, large-neighborhood explorations based on mixed-integer linear programming techniques are now available to diversify the search and to escape local optima related to small neighborhoods, to ultimately find better solutions. It consists essentially in exploiting the linear relaxation of appropriate subproblems, coupled with primal rounding heuristics to find quickly good-quality integer solutions. In this way, when neighborhoods (that is, subproblems) have a good linear relaxation, LocalSolver is now able to exploit it. The linear relaxation is computed using a revised dual simplex algorithm.
Minor features:
- LocalSolver now automatically manages impossible values for some operators: zero value for division, negative values for square root, zero value for logarithm, etc. With such impossible values in input, the operator outputs Not-A-Number (NaN) as value. LocalSolver automatically searches for solutions not leading to NaN values for these operators.
- LocalSolver now automatically manages an empty list of operands for n-ary operators like: sum, prod, or, and, xor. The value returned by an empty expression corresponds to the neutral value for the corresponding operator. For example: for sum the neutral value is 0; for prod it is 1; for or it is false or 0; for and it is true or 1; for xor it is false or 0.
Among the classes of problems you can newly tackle with LocalSolver are: supply chain optimization problems, unit commitment problems, portfolio optimization problems, as well as diverse numerical optimization problems. The main advantage of using LocalSolver over other mathematical optimization solvers of the market is to make you able to tackle ultra-large problems involving millions of variables within short computation times.
LocalSolver is free for academics: create your account in one minute on http://www.localsolver.com/account.html to try it!