Distributed Simulated Annealing Muhammad Arshad * and Marius C. Silaghi Florida Institute of Technology {marshad,msilaghi}@cs.fit.edu Abstract. Distributed Constraint Satisfaction is a framework for modeling and solving distributed problems. Research on the topic intensified during the last ten years when mainly systematic techniques were thoroughly explored. A revival of attention for Distributed Stochastic Algorithms (DSA) was marked by the work of [3, 4, 13, 7, 5]. Remarkably, it was proven that Distributed Stochastic Search is a competitive tech- nique. Here we review current DSAs and their intrinsic properties. We also describe a remarkable new algorithm, Distributed Simulated Annealing (DSAN). We show theo- retically and experimentally how it compares with DSA. Experimental evaluation on synchronous versions shows that with current heuristics DSAN competes with the pre- vious winners, DSA- B and DSA- C. Depending on the parameters selected for DSA, DSAN may offer marginally better quality solutions than DSA- B or DSA- C for hard and over-constrained problems. For easy problems DSAN may not lock on the solution and use of termination detection is required. 1 Introduction Constraint satisfaction has proven to be a successful paradigm for approaching combinatorial problems like resource allocation, scheduling, or planning in centralized settings. A constraint satisfaction problem (CSP) is given by: • a set of variables {x 1 ,x 2 , ..., x n }, • a set of domains, {D 1 ,D 2 , ..., D n }, associated with the variables, and • a set of constraints, {C 1 ,C 2 , ..., C k }, each of them involving a subset of the set of vari- ables, The solution to a CSP is an assignment of values from the corresponding domains to each variable such that the obtained combination is allowed by each constraint. A distributed CSP (DCSP) arises when variables are distributed among agents so that each variable can only be assigned values by a single agent [12]. This is the definition exploited in our technique, even if DSA/DSAN can be easily extended to other frameworks of distributed CSPs, notably where assignments for each variable can be proposed by several agents [8]. Distributed Constraint Satisfaction can model and solve naturally distributed problems. Research on DCSPs has accelerated during the last years when complete techniques as well as techniques based on tight organization were thoroughly explored. Attention was drawn to Distributed Stochastic Algorithms (DSA) by the work of [3, 4, 13, 7, 5]. Remarkably, it was * Undergraduate student at FIT.