## Reactive Search Optimization: Learning while Optimizing

Citations: | 4 - 2 self |

### BibTeX

@MISC{Battiti_reactivesearch,

author = {Roberto Battiti and Mauro Brunato},

title = {Reactive Search Optimization: Learning while Optimizing},

year = {}

}

### OpenURL

### Abstract

The final purpose of Reactive Search Optimization (RSO) is to simplify the life for the final user of optimization. While researchers enjoy designing algorithms, testing alternatives, tuning parameters and choosing solution schemes — in fact this is part of their daily life — the final users ’ interests are different: solving a problem in the

### Citations

3517 | Optimization by simulated annealing
- KIRKPATRICK, GELATT, et al.
- 1983
(Show Context)
Citation Context ...e configuration variables and optimizing subparts of the problem [92]. 2.3 Reacting on the annealing schedule A widely popular stochastic local search technique is the Simulated Annealing (SA) method =-=[88]-=- based on the theory of Markov processes. The trajectory is built in a randomized manner: the successor of the current point is chosen stochastically, with a probability that depends only on the diffe... |

1043 |
An efficient heuristic procedure for partitioning graphs
- Kernighan, Lin
- 1970
(Show Context)
Citation Context ...g connectivity by means of homogeneous connections. The graph partitioning problem has been a test case for advanced local search heuristics starting at least from the seminal Kernighan and Lin paper =-=[84]-=-, which proposes a variable-depth schemes. This is is fact a simple prohibition-based (tabu) scheme where swaps of nodes among the two sets of the partitions are applied, and the just swapped nodes ar... |

703 |
Data Structures and Algorithms
- Aho, Hopcroft, et al.
- 1985
(Show Context)
Citation Context ...lued function ALLOW selects a subset of N(X (t+1) ) in a manner that depends on the entire search trajectory X (0) ,...,X (t+1) . By analogy with the concept of abstract data type in Computer Science =-=[2]-=-, and with the related object-oriented software engineering framework [49], it is useful to separate the abstract concepts and operations of TS from the detailed implementation, i.e., realization with... |

680 | A New Method for Solving Hard Satisfiability Problems
- Selman, Levesque, et al.
- 1992
(Show Context)
Citation Context ...terals, a literal is a variable or its negation). The decision version is called SAT, one searches for a variable assignment, if any exists, which makes a formula true. The influential algorithm GSAT =-=[117]-=- is based on local search with the standard basic moves flipping the individual variables (from false to true and vice versa). Different noise strategies to escape from locally optimal configurations ... |

471 |
Tabu Search-Part I
- Glover
- 1989
(Show Context)
Citation Context ...ased heuristics” starting from the late eighties is greatly due to the role of F. Glover in the proposal and diffusion of a rich variety of meta-heuristic tools under the umbrella of Tabu Search (TS) =-=[68, 69]-=-, but see also [73] for an independent seminal paper. It is evident that Glover’s ideas have been a source of inspiration for many approaches based on the intelligent use of memory in heuristics. The ... |

361 | Noise strategies for improving local search
- Selman, Kautz, et al.
- 1994
(Show Context)
Citation Context ...arch with the standard basic moves flipping the individual variables (from false to true and vice versa). Different noise strategies to escape from locally optimal configurations are added to GSAT in =-=[115, 116]-=-. In particular, the GSAT-with-walk algorithm introduces random walk moves with a certain probability. A prototypical evaluation function modification algorithm is the breakout method proposed in [98]... |

271 |
Tabu Search: Part II
- Glover
- 1990
(Show Context)
Citation Context ...ased heuristics” starting from the late eighties is greatly due to the role of F. Glover in the proposal and diffusion of a rich variety of meta-heuristic tools under the umbrella of Tabu Search (TS) =-=[68, 69]-=-, but see also [73] for an independent seminal paper. It is evident that Glover’s ideas have been a source of inspiration for many approaches based on the intelligent use of memory in heuristics. The ... |

269 | B.: Local search strategies for satisfiability testing
- Selman, Kautz, et al.
- 1996
(Show Context)
Citation Context ...arch with the standard basic moves flipping the individual variables (from false to true and vice versa). Different noise strategies to escape from locally optimal configurations are added to GSAT in =-=[115, 116]-=-. In particular, the GSAT-with-walk algorithm introduces random walk moves with a certain probability. A prototypical evaluation function modification algorithm is the breakout method proposed in [98]... |

223 | The Reactive Tabu Search
- Battiti, Tecchiolli
- 1994
(Show Context)
Citation Context ... −1 ) < (t − T (t) )} (5) X (t+1) = BEST-NEIGHBOR(NA(X (t) )) (6) Rules to determine the prohibition parameter by reacting to the repetition of previously-visited configurations have been proposed in =-=[26]-=- (reactive-TS, RTS for short). In addition, there are situations where the single reactive mechanism on T is not sufficient to avoid long cycles in the search trajectory and therefore a second reactio... |

215 | Domain-independent extensions to GSAT: Solving large structured satisfiability problems
- Selman, Kautz
- 1993
(Show Context)
Citation Context ...m a given attraction basin. New clause-weighing parameters are introduced and therefore new possibilities for tuning the parameters based on feedback from preliminary search results. The algorithm in =-=[113]-=- suggests to use weights to encourage more priority on satisfying the “most difficult” clauses. One aims at learning how difficult a clause is to satisfy. These hard clauses are identified as the ones... |

205 | Variable neighborhood search
- Mladenović, Hansen
- 1997
(Show Context)
Citation Context ...ctory goes deeper and deeper into a given local minimum attractor. When the neighborhood is changed depending on the local configuration one obtains for example the Variable Neighborhood Search (VNS) =-=[72]-=-. VNS considers a a set of neighborhoods, defined a priori at the beginning of the search, and then uses the most appropriate one during the search. Variable Neighborhood Descent [74] (VND), see Fig. ... |

194 |
The breakout method for escaping from local minima
- Morris
- 1993
(Show Context)
Citation Context ...116]. In particular, the GSAT-with-walk algorithm introduces random walk moves with a certain probability. A prototypical evaluation function modification algorithm is the breakout method proposed in =-=[98]-=- for the related constraint satisfaction problem. The cost is measured as the sum of the weights associated to the violated constraints. Each weight is one at the beginning, at a local minimum the wei... |

181 | Very fast simulated re-annealing
- Ingber
- 1989
(Show Context)
Citation Context ...f the minimization process” can deliver precious feedback about some crucial internal parameters of the algorithm. In Adaptive Simulated Annealing (ASA), also known as very fast simulated reannealing =-=[82]-=-, the parameters that control the temperature cooling schedule and the random step selection are automatically adjusted according to algorithm progress. If the state is represented as a point in a box... |

164 |
Minirnizing multimodal functions of continuous variables with the 'Simulated Annealing' algorithm
- Corana, Marchesi, et al.
- 1987
(Show Context)
Citation Context ...to the optimum value.Reactive Search Optimization: Learning while Optimizing 13 The application of SA to continuous optimization (optimization of functions defined on real variables) is pioneered by =-=[48]-=-. The basic method is to generate a new point with a random step along a direction eh, to evaluate the function and to accept the move with the exponential acceptance rule. One cycles over the differe... |

136 | Towards an understanding of hill-climbing procedures for SAT
- Gent, Walsh
- 1993
(Show Context)
Citation Context ...ochastic sampling: only a partial list of candidates is examined before choosing the best allowed neighbor. Finally, other possibilities which are softer than prohibitions exist. For example the HSAT =-=[67]-=- variation of GSAT introduces a tie-breaking rule into GSAT: if more moves produce the same (best) ∆ f , the preferred move is the one that has not been applied for the longest span. HSAT can be seen ... |

132 |
Robust taboo search for the quadratic assignment problem
- Taillard
- 1991
(Show Context)
Citation Context ...mptotic results for TS can be obtained in probabilistic TS [56]. In a different proposal (robust-TS) the prohibition parameter is randomly changed between an upper and a lower bound during the search =-=[122]-=-. If the neighborhood evaluation is expensive, the exhaustive evaluation can be substituted with a partial stochastic sampling: only a partial list of candidates is examined before choosing the best a... |

111 |
Metastrategy Simulated Annealing and Tabu Search Algorithms for the Vehicle Routing Problem
- Osman
- 1993
(Show Context)
Citation Context ...hm, so that longer allocated CPU times are related to possibly better and better values until the user decides to stop. Non-monotonic cooling schedules are a reactive solution to this difficulty, see =-=[46, 105, 1]-=-. The work [46] suggests to reset the temperature once and for all at a constant temperature high enough to escape local minima but also low enough to visit them, for example, at the temperature Tfoun... |

106 |
Algorithms for the maximum satisfiability problem
- Hansen, Jaumard
- 1990
(Show Context)
Citation Context ...ng from the late eighties is greatly due to the role of F. Glover in the proposal and diffusion of a rich variety of meta-heuristic tools under the umbrella of Tabu Search (TS) [68, 69], but see also =-=[73]-=- for an independent seminal paper. It is evident that Glover’s ideas have been a source of inspiration for many approaches based on the intelligent use of memory in heuristics. The main competitive ad... |

104 |
Object oriented programming: an evolutionary approach
- Cox
- 1986
(Show Context)
Citation Context ...nds on the entire search trajectory X (0) ,...,X (t+1) . By analogy with the concept of abstract data type in Computer Science [2], and with the related object-oriented software engineering framework =-=[49]-=-, it is useful to separate the abstract concepts and operations of TS from the detailed implementation, i.e., realization with specific data structures. In other words, policies (that determine which ... |

99 | An empirical study of greedy local search for satisfiability testing
- Selman, Kautz
- 1993
(Show Context)
Citation Context ...ght is one at the beginning, at a local minimum the weight of each nogood is increased by one until one escapes from the given local minimum (a breakout occurs). Clause-weighting has been proposed in =-=[114]-=- for GSAT. A positive weight is associated to each clause to determine how often the clause should be counted when determining which variable to flip. The weights are dynamically modified during probl... |

92 | Large-step markov chains for the traveling salesman problem
- Martin, Otto, et al.
- 1991
(Show Context)
Citation Context ...building a sequence of locally optimal solutions by perturbing the current local minimum and applying local search after starting from the modified solution. The work about large-step Markov chain of =-=[96, 94, 95, 126]-=- contains very interesting results coupled with a clear description of the principles. In VNS minimal perturbations maintain the trajectory in the starting attraction basin, while excessive ones bring... |

90 |
Simulated Annealing: Theory and Applications
- L
- 1988
(Show Context)
Citation Context ...while the minimum change in energy defines the minimum-temperature scale. These temperature scales tell us where to begin and end an annealing schedule. The analogy with physics is further pursued in =-=[89]-=-, where concepts related to phase transitions and specific heat are used. The idea is that a phase transition is related to solving a sub-part of a problem. After a phase transition corresponding to1... |

87 | Scaling and probabilistic smoothing: Efficient dynamic local search for SAT
- Hutter, Tompkins, et al.
- 2002
(Show Context)
Citation Context ...ts which are very far from the local minimum, but which share some of the unsatisfied clauses, will also see their values changed. A more recent proposal of a dynamic local search (DLS) for SAT is in =-=[124]-=-. The authors start from the Exponentiated Sub-Gradient (ESG) algorithm [112], which alternates search phases and weight updates, and develop a scheme with low time complexity of its search steps: Sca... |

80 | Combining simulated annealing with local search heuristics
- Martin, Otto
- 1996
(Show Context)
Citation Context ...building a sequence of locally optimal solutions by perturbing the current local minimum and applying local search after starting from the modified solution. The work about large-step Markov chain of =-=[96, 94, 95, 126]-=- contains very interesting results coupled with a clear description of the principles. In VNS minimal perturbations maintain the trajectory in the starting attraction basin, while excessive ones bring... |

75 | Reactive local search for the maximum clique problem, Algorithmica 29 (4
- Battiti, Protasi
- 2001
(Show Context)
Citation Context ... Problem with Time Windows is proposed in [35]. Vehicle routing with soft time windows and Erlang travel times is studied in [109]. An RSO scheme is applied to the maximum clique problem in graphs in =-=[19]-=- [23]. A clique is a subset of nodes which are mutually interconnected, the problem is related to identifying densely interconnected communities and, in general, to clustering issues. A relaxed quasi-... |

69 | Increasing Internet capacity using local search
- Fortz, Thorup
(Show Context)
Citation Context ...ication medium. In such a network, a hidden group may try to camouflage its communications amongst the typical communications of the network. The task of increasing internet capacity is considered in =-=[62]-=-. The multiple-choice multi-dimensional knapsack problem (MMKP) with applications to service level agreements and multimedia distribution is studied in [76] and [77]. They consider the model of alloca... |

65 |
An improved Annealing scheme for the QAP
- Connolly
- 1990
(Show Context)
Citation Context ...hm, so that longer allocated CPU times are related to possibly better and better values until the user decides to stop. Non-monotonic cooling schedules are a reactive solution to this difficulty, see =-=[46, 105, 1]-=-. The work [46] suggests to reset the temperature once and for all at a constant temperature high enough to escape local minima but also low enough to visit them, for example, at the temperature Tfoun... |

59 | Learning evaluation functions for global optimization and Boolean satisfiability
- Boyan, Moore
- 1998
(Show Context)
Citation Context ... a dynamically modified (learned) evaluation function is related to the machine learning technique of reinforcement learning (RL). Early applications of RL in the area of local search is presented in =-=[34, 33]-=-. Some reinforcement learning approaches for optimization are also discussed in [8]. Recent work includes [15], on-the-fly parameter tuning for evolutionary algorithms in [55], and the presentation in... |

59 | Automatic algorithm configuration based on local search
- Hutter, Hoos, et al.
- 2007
(Show Context)
Citation Context ...nce basis by simply picking the parameter configuration that is predicted to yield the lowest run-time. An iterated local search (ILS) algorithm for the algorithm configuration problem is proposed in =-=[81]-=-. The approach works for both deterministic and randomized algorithms and can be applied regardless of tuning scenario and optimization objective. On-line and off-line strategies are complementary: in... |

58 | Large-step markov chains for the tsp incorporating local search heuristics
- Martin, Otto, et al.
- 1992
(Show Context)
Citation Context ...building a sequence of locally optimal solutions by perturbing the current local minimum and applying local search after starting from the modified solution. The work about large-step Markov chain of =-=[96, 94, 95, 126]-=- contains very interesting results coupled with a clear description of the principles. In VNS minimal perturbations maintain the trajectory in the starting attraction basin, while excessive ones bring... |

56 | Learning evaluation functions to improve optimization by local search
- Boyan, Moore
- 2000
(Show Context)
Citation Context ... a dynamically modified (learned) evaluation function is related to the machine learning technique of reinforcement learning (RL). Early applications of RL in the area of local search is presented in =-=[34, 33]-=-. Some reinforcement learning approaches for optimization are also discussed in [8]. Recent work includes [15], on-the-fly parameter tuning for evolutionary algorithms in [55], and the presentation in... |

54 | Performance prediction and automated tuning of randomized and parametric algorithms
- Hutter, Hamadi, et al.
- 2006
(Show Context)
Citation Context ...timization, intelligent tuning and design of heuristics. The RSO approach of learning on the job is to be contrasted with off-line parameter tuning. This orthogonal approach is studied for example in =-=[80, 79]-=-, that proposes methods to predict per-instance and per-parameter run-times with reasonable accuracy. These predictive models are then used to predict which parameter settings result in the lowest run... |

50 |
Concepts of scale in simulated annealing
- White
- 1984
(Show Context)
Citation Context ...rting temperature Tstart, geometric cooling schedule Tt+1 = α Tt, with α < 1, final temperature Tend), a sensible choice has to be made for the three involved parameters Tstart, α, and Tend. The work =-=[130]-=- suggests to estimate the distribution of f values. The standard deviation of the energy distribution defines the maximum-temperature scale, while the minimum change in energy defines the minimum-temp... |

49 | Guided Local Search and its application to the travelling salesman problem
- Voudouris, Tsang
- 1999
(Show Context)
Citation Context ...uced that adaptively tunes one of the algorithm’s important parameters. A similar approach of dynamically modifying the objective function has been proposed with the term of Guided Local Search (GLS) =-=[127, 128]-=- for other applications. GLS aims at enabling intelligent search schemes that exploit problem- and search-related information to guide a local search algorithm. Penalties depending on solution feature... |

46 | Boosting verification by automatic tuning of decision procedures
- Hutter, Babić, et al.
- 2007
(Show Context)
Citation Context ...timization, intelligent tuning and design of heuristics. The RSO approach of learning on the job is to be contrasted with off-line parameter tuning. This orthogonal approach is studied for example in =-=[80, 79]-=-, that proposes methods to predict per-instance and per-parameter run-times with reasonable accuracy. These predictive models are then used to predict which parameter settings result in the lowest run... |

44 | The exponentiated subgradient algorithm for heuristic boolean programming
- Schuurmans, Southey, et al.
- 2001
(Show Context)
Citation Context ...atisfied clauses, will also see their values changed. A more recent proposal of a dynamic local search (DLS) for SAT is in [124]. The authors start from the Exponentiated Sub-Gradient (ESG) algorithm =-=[112]-=-, which alternates search phases and weight updates, and develop a scheme with low time complexity of its search steps: Scaling and Probabilistic Smoothing (SAPS). Weights of satisfied clauses are mul... |

42 |
A tabu search approach to the constraint satisfaction problem as a general problem solver
- Nonobe, Ibaraki
- 1998
(Show Context)
Citation Context ...ysensitive) scheme.18 Roberto Battiti and Mauro Brunato Maximum satisfiability is considered in [21], [20] [97], reactive SAPS (scaling and probabilistic smoothing) [124]. Constraint satisfaction in =-=[102]-=-. Reactive local search techniques for the maximum k-conjunctive constraint satisfaction problem (MAX-k-CCSP) in [22]. A worst-case analysis of tabu search as a function of the tabu list length for th... |

40 |
Simulated Annealing and Combinatorial Optimization
- Nahar, Sahni, et al.
- 1986
(Show Context)
Citation Context ..., and in variable neighborhood search. Modifications departing from the exponential acceptance rule and other adaptive stochastic local search methods for combinatorial optimization are considered in =-=[99, 100]-=-. The authors appropriately note that the optimal choices of algorithm parameters depend not only on the problem but also on the particular instance and that a proof of convergence to a globally optim... |

34 |
Solving the Pickup and Delivery Problem with Time Windows Using Reactive Tabu Search, Transportation Research Part B
- Nanry, Barnes
- 2000
(Show Context)
Citation Context ...ess. The adaptive memory strategy takes the search back to the unexplored regions of the search space by maintaining a set of elite solutions and using them strategically with the RTS. The authors of =-=[101]-=- address the pickup and delivery problem with time windows using reactive tabu search. Real-time dispatch of trams in storage yards is studied in [131]. In the military sector, simple versions of Reac... |

33 | Learning short-term weights for GSAT
- Frank
- 1997
(Show Context)
Citation Context ...scent followed by plateau search. Their weight is increased so that future runs will give them more priority when picking a move. More algorithms based on the same weighting principle are proposed in =-=[63, 64]-=-, where clause weights are updated after each flip: the reaction from the unsatisfied clauses is now immediate as one does not wait until the end of a try (weighted GSAT or WGSAT). If weights are only... |

32 | Training Neural Nets with the Reactive Tabu Search
- Battiti, Tecchiolli
(Show Context)
Citation Context ... the presence of local minima. In addition, the calculation of derivatives is expensive and error-prone, especially if special-purpose VLSI hardware is used. A radically different approach is used in =-=[29]-=-: the task is transformed into a combinatorial optimization problem (the points of the search space are binary strings), and solved with the Reactive Search Optimization algorithm. To speed up the nei... |

31 |
Reactive search, a history-sensitive heuristic for MAXSAT
- Battiti, Protasi
- 1997
(Show Context)
Citation Context ... appropriate distance measure from a given starting configuration. An insufficient growth if the distance as a function of the number of steps can be taken as evidence of confinement, see for example =-=[20]-=-. A more radical escape mechanism can be triggered when the basic prohibition mechanism is not sufficient to guarantee diversification. In [26] the escape (a number of random steps) is triggered when ... |

31 | Partial constraint satisfaction problems and guided local search, in
- Voudouris, Tsang
- 1996
(Show Context)
Citation Context ...uced that adaptively tunes one of the algorithm’s important parameters. A similar approach of dynamically modifying the objective function has been proposed with the term of Guided Local Search (GLS) =-=[127, 128]-=- for other applications. GLS aims at enabling intelligent search schemes that exploit problem- and search-related information to guide a local search algorithm. Penalties depending on solution feature... |

30 | Greedy, prohibition, and reactive heuristics for graph partitioning
- Battiti, Bertossi
- 1999
(Show Context)
Citation Context ...t the search trajectory must go before it is allowed to come back to a previously visited point. In particular, the following relationships between prohibition and diversification are demonstrated in =-=[11]-=- for a search space consisting of binary strings with basic moves flipping individual bits: • The Hamming distance H between a starting point and successive point along the trajectory is strictly incr... |

30 | Weighting for Godot: Learning Heuristics for Gsat
- Frank
- 1996
(Show Context)
Citation Context ...scent followed by plateau search. Their weight is increased so that future runs will give them more priority when picking a move. More algorithms based on the same weighting principle are proposed in =-=[63, 64]-=-, where clause weights are updated after each flip: the reaction from the unsatisfied clauses is now immediate as one does not wait until the end of a try (weighted GSAT or WGSAT). If weights are only... |

26 |
Simulated Annealing Cooling Schedules for the School Timetabling Problem”, Asia-Pacific
- Abramson, Dang, et al.
- 1999
(Show Context)
Citation Context ...hm, so that longer allocated CPU times are related to possibly better and better values until the user decides to stop. Non-monotonic cooling schedules are a reactive solution to this difficulty, see =-=[46, 105, 1]-=-. The work [46] suggests to reset the temperature once and for all at a constant temperature high enough to escape local minima but also low enough to visit them, for example, at the temperature Tfoun... |

26 |
Tabu Search applied to global optimization
- Chelouah, Siarry
- 2000
(Show Context)
Citation Context ... be optimized ( boxes are larger in “flat” regions, smaller in regions with a “rough” structure). An application of intelligent prohibition-based strategies to continuous optimization is presented in =-=[43]-=-. A Reactive Affine Shaker method for continuous optimization is studied in [36, 37]. The work presents an adaptive stochastic search algorithm for the optimization of functions of continuous variable... |

22 | Efficient parameter estimation for RNA secondary structure prediction, Bioinformatics 23
- Andronescu, Condon, et al.
- 2007
(Show Context)
Citation Context ...tions provided by the system. The preference a � b of the DM for the solution a w.r.t. the solution b is represented by the following constraint: g(a,w) < g(b,w) Our approach, inspired by the work in =-=[1]-=-, consists of finding a solution w for the set of constraints on the weights generated by interacting with the DM. All the weights are initialized with random values in the interval (0,1), and then no... |

21 | A reactive variable neighborhood search for the vehicle-routing problem with time windows
- Braysy
- 2003
(Show Context)
Citation Context ...is activated when there is evidence of entrapment in an attraction basin and gradually reduced when there is evidence that a new basin has been discovered. An explicitly reactive-VNS is considered in =-=[35]-=- for the VRPTW problem (vehicle routing problem with time windows), where a construction heuristic is combined with VND using first-improvement local search. Furthermore, the objective function used b... |

21 |
A Reactive Tabu Search Metaheuristic for the Vehicle Routing Problem with Time Windows
- Chiang, Russell
- 1997
(Show Context)
Citation Context ... based on producing coarse versions of very large graphs are considered for Graph Partitioning in [10]. A reactive tabu search version for the vehicle routing problem with time windows is designed in =-=[44]-=-, while a version of the vehicle routing problem with back-hauls is considered in[50] and [104]. A Reactive Variable Neighborhood Search for the Vehicle-Routing Problem with Time Windows is proposed i... |