4.1. Ant colony optimization
The ACO algorithm mimics the collective behavior of ant foraging, which was firstly introduced by Colorni et al. (1991). During the searching process for food sources, ants behave intelligently to find the optimal path to food source, which is practically achieved by the utilization of pheromone. The existence of pheromone shows the trace of an ant, and provides heuristic information for other ants, which decide whether to follow this pheromone trace or not. If the new ant chooses to follow this pheromone trace, it would reinforce the density of pheromone; otherwise the pheromone would be gradually evaporated and finally exhausted. The above decision strategies can be regarded as positive feedback and negative feedback respectively. Higher pheromone density indicates higher chosen probability. Therefore, more and more ants choose to follow the trace with high pheromone density and construct the optimal path to the food source. Since the introduction of the ACO algorithm, it has gained much popularity and diverse variants of ACO have also been proposed for adapting various CO problems, such as Ant System (Dorigo et al., 1996), Ant Colony System (Dorigo and Gambardella, 1997a) and MAX–MIN Ant System (Stützle and Hoos, 2000). More comprehensive descriptions of ant related algorithms can be found in the literature (Cordon et al., 2002).
The whole procedure of the ACO algorithm can be illustrated as follows. Initially, ants are placed on the nodes of the graph randomly. Then each ant selects a connected and unvisited node as its next movement probabilistically. The probability is influenced two factors, the distance from the current node to the next node and the pheromone on the associated edge. This movement is executed iteratively until all ants have traversed all the nodes on the graph, which is called one cycle. After each cycle, the pheromone deployment of the whole graph is updated. The principle is that whenever an ant moves through an edge, the pheromone on that edge is reinforced; otherwise, it would evaporate and be exhausted. After a certain number of cycles, the path with highest pheromone density is found, which represents the optimal solution.
The applications of ACO in green logistics are of great value as the intrinsic features of the ACO algorithm, finding the shortest route by ants, promote its popular application in vehicle routing problems. For example,Dorigo and Gambardella (1997a) firstly applied ant colonies to solve the standard TSP, while Bell and McMullen (2004) further adapted ACO techniques to handle the VRP. Moreover, Moncayo-Martínez and Zhang (2011) adopted a multi-objective ant colony optimization approach to resolve the supply chain design problem.
Apart from the general adoption of ACO, researchers are also trying to improve the effectiveness and efficiency of ACO through various means. Chen and Ting (2006) proposed an improved ant colony system (IACS) algorithm, possessing a new state transition rule, a new pheromone update rule and diverse local search approaches, to solve VRP. Agrawal and Tiwari (2007) introduced collaborative ant colony optimization (CACO) to resolve the stochastic mixed-model U-shaped disassembly line balancing and sequencing problem. The distinguishing feature of this proposed CACO was that it maintained bilateral colonies of ants which independently identified the two disassembly sequences, but using the information obtained by their collaboration to guide to the future path. Gajpal and Abad (2009b) presented a multi-ant colony system (MACS) using a new construction rule as well as two multi-route local search schemes for VRP with backhauls. Yu et al. (2009) brought forward an improved ant colony optimization (IACO) including a new pheromone updating strategy, called ant-weight strategy, and a mutation operation. Çatay (2010) came up a saving-based ant algorithm by employing a new saving-based visibility function and pheromone updating procedure. Liu et al. (2012) improved the max–min ant system based on the strategy of sorting elite ants to tackle the disassembly sequence planning.
Another direction for the improvement of the ACO algorithm is the hybridization with other techniques. Other techniques could either provide good initial solutions as the input for ACO at the beginning, or perform the function of improving the middle candidate solutions. For instance, Chen and Ting (2008) combined the Lagrangian heuristic and the ant colony system to create a new hybrid algorithm, named LH-ACS, to handle the single source capacitated FLP. Lee et al. (2010) proposed an enhanced ant colony optimization (EACO), in which the simulated annealing (SA) provided initial solutions for ACO. Zhu and Zhang (2010) presented ant colony optimization with elitist ant (ACOEA) algorithm. Both tabu search (TS) and elitist strategy were adopted in this ACOEA to improve the performance of candidate solutions. Balseiro et al. (2011) hybridized ACO with an aggressive insertion heuristic to overcome the shortcoming of ACO in case of infeasible solutions in VRP. Wang (2013) presented an adaptive ant colony algorithm (AACA), coupled with a pareto local search algorithm, to overcome the premature convergence problem when applying ACO into VRP.
4.2. Particle swarm optimization
The PSO algorithm simulates the movement of a set of particles in search space under predetermined rules in order to find the optimal position. PSO was originally proposed by Eberhart and Kennedy (1995), being inspired by bird flocking, fish schooling, and even human social behavior. Reynolds (1987) proposed a bird-oid model to simulate the behavior of bird flocking. Therein each individual followed three simple rules: collision avoidance, velocity matching, and flock centering. Derived from the bird-oid model, in PSO, individuals representing solutions herein are treated as particles, and each particle is characterized by its associated fitness value, position vector and velocity vector. Apart from the three inner attributes, each particle also memorizes its historical best position (local best position) and global best position of the swarm, and refers to those two positions whenever it moves to the next position. During the iterative process of movement, all particles gradually converge at the global optimal position.
The procedure of the PSO algorithm can be described as in the following. At the beginning, a number of particles are randomly placed in the search space. Each particle holds its position and velocity information in a vector format. Whenever movement occurs, the particle needs to update its velocity information firstly, referring to three factors: its current velocity, the local best position and the global best position. Different weightings of different factors indicate different optimization strategies. Subsequently, the particle updates its position information following the updated velocity vector. The positions of each particle correspond to candidate solutions. The local and global best positions are updated after each movement provided that the particle arrives at a better position. This procedure is conducted iteratively until the stopping criteria are met. The global best position is the optimal solution which can be found so far.
Wang and Liu (2010) proposed a chaotic particle swarm optimization (CPSO) approach to handle the assembly line balancing problem, in which the chaos method was utilized to improve the solution quality and to increase the convergence rate. Kanthavel et al. (2012) developed nested particle swarm optimization, as the integration of two other mechanism, master particle swarm optimization (MPSO) and slave particle swarm optimization (SPSO), to tackle the VRP with simultaneously delivery and pickup. Shankar et al. (2012) introduced a hybrid multi-objective particle swarm optimization (MOPSO) algorithm to solve the bi-objective distribution scheduling, while Venkatesan and Kumanan (2012) employed another multi-objective discrete particle swarm algorithm (MODPSA), containing two different global guide selection techniques, for supply chain network design.
Qi (2011) proposed an improved discrete particle swarm optimization (DPSO), in which an iterated local search (ILS) method was adopted to ensure the avoidance of the local minimum. Latha Shankar et al. (2013)integrated a non-dominated sorting (NDS) procedure in a multi-objective hybrid particle swarm optimization algorithm (MOHPSO) to achieve bi-objective optimization of two conflicting objectives. Validi et al. (2013)presented a design of an experiment (DoE) guided multiple-objective particle swarm optimization (MOPSO) optimizer, in which DoE was utilized to eliminate the un-realistic set of feasible and optimal solution sets, while another popular multi-attribute decision-making approach, TOPSIS, was employed to evaluate the solutions through exhaustive analysis, e.g. prioritization, ranking and scenario analysis.
4.3. Artificial bee colony
The artificial bee colony (ABC) algorithm was introduced by Karaboga (2005), which imitates the collective and collaborative forging behavior of a bee colony. When forging, different bees work collaboratively to explore and exploit food sources with rich nectar. The artificial bee colony consists of three types of bees, scout bees, employed bees and onlooker bees, which play different roles in the exploration and exploitation of food sources. Food sources are regarded as the solutions of specific problems, among which the ones with more nectar correspond to better solutions.
The whole procedure of the ABC algorithm can be described as follows. Scout bees are assigned to find the initial food sources by carrying out a random search in the search space. After that, employed bees are sent out to exploit the discovered food sources, and each employed bee matches one food source. During the exploitation procedure, each employed bee also carries out a neighb