Metaheuristics are generally applied to problems for which there is no satisfactory problem-specific algorithm or heuristic; or when it is not practical to implement such a method. Most commonly used metaheuristics are targeted to combinatorial optimization problems, but of course can handle any problem that can be recast in that form, such as solving boolean equations.
The function to be optimized is called the goal function, or objective function, and is usually provided by the user as a black-box procedure that evaluates the function on a given state. Depending on the meta-heuristic, the user may have to provide other black-box procedures that, say, produce a new random state, produce variants of a given state, pick one state among several, provide upper or lower bounds for the goal function over a set of states, and the like.
Some metaheuristics maintain at any instant a single current state, and replace that state by a new one. This basic step is sometimes called a state transition or move. The move is uphill or downhill depending on whether the goal function value increases or decreases. The new state may be constructed from scratch by a user-given generator procedure. Alternatively, the new state be derived from the current state by a user-given mutator procedure; in this case the new state is called a neighbour of the current one. Generators and mutators are often probabilistic procedures. The set of new states that can be produced by the mutator is the neighbourhood of the current state.
More sophisticated meta-heuristics maintain, instead of a single current state, a current pool with several candidate states. The basic step then may add or delete states from this pool. User-given procedures may be called to select the states to be discarded, and to generate the new ones to be added. The latter may be generated by combination or crossover of two or more states from the pool.
A metaheuristic also keep track of the current optimum, the optimum state among those already evaluated so far.
Since the set of candidates is usually very large, metaheuristics are typically implemented so that they can be interrupted after a client-specified time budget. If not interrupted, some exact metaheuristics will eventually check all candidates, and use heuristic methods only to choose the order of enumeration; therefore, they will always find the true optimum, if their time budget is large enough. Other metaheuristics give only a weaker probabilistic guarantee, namely that, as the time budget goes to infinity, the probability of checking every candidate tends to 1.
Innumerable variants and hybrids of these techniques have been proposed, and many more applications of metaheuristics to specific problems have been reported. This is an active field of research, with a considerable literature, a large community of researchers and users, and a wide range of applications.
Those critics point out, for one thing, that the general goal of the typical metaheuristic — the efficient optimization of an arbitrary black-box function—cannot be solved efficiently, since for any metaheuristic M one can easily build a function f that will force M to enumerate the whole search space (or worse). Indeed, the "no-free-lunch theorem" says that over the set of all mathematically possible problems, each optimization algorithm will do on average as well as any other. Thus, at best, a specific metaheuristic can be efficient only for restricted classes of goal functions (usually those that are partially "smooth" in some sense). However, when these restrictions are stated at all, they either exclude most applications of interest, or make the problem amenable to specific solution methods that are much more efficient than the meta-heuristic.
Moreover, all metaheuristics rely on auxiliary procedures (producers, mutators, etc.) that are given by the user as black-box functions. It turns out that the effectiveness of a metaheuristic on a particular problem depends almost exclusively on these auxiliary functions, and very little on the metaheuristic itself. Given any two distinct metaheuristics M and N, and almost any goal function f, it is usually possible to write a set of auxiliary procedures that will make M find the optimum much more efficient than N, by many orders of magnitude; or vice-versa. In fact, since the auxiliary procedures are usually unrestricted, one can submit the basic step of metaheuristic M as the generator or mutator for N. Because of this extreme generality, one cannot say that any metaheuristic is better than any other, not even for a specific class of problems. In particular, no meta-heuristic can be shown to be better for any specific problem than brute force search, or the following "banal metaheuristic":
Finally, all metaheuristic optimization techniques are extremely crude when evaluated by the standards of (continuous) nonlinear optimization. Within this area, it is well-known that to find the optimum of a smooth function on n variables one must essentially obtain its Hessian matrix, the n by n matrix of its second derivatives. If the function is given as a black-box procedure, then one must call it about n2/2 times, and solve an n by n system of linear equations, before one can make the first useful step towards the minimum. However, none of the common metaheuristics incorporate or accommodate this procedure. At best, they can be seen as computing some crude approximation to the local gradient of the goal function, and moving more or less "downhill". But gradient-descent is can be extremely inefficient for non-linear optimization. For example, consider the problem of finding a pair of numbers x,y that minimizes the quadratic function Q(x,y) = 1000000(x + y - 1000)2 + (x - y - 10)2. Gradient-descent methods will generally take a very long time to reach the minimum from, say, (1000,0); whereas Hessian-based methods will reach it in one step. Unfortunately, "narrow valley" functions like this one are increasingly likely to occur as the dimension of the space increases.
Even though meta-heuristics are often used for discrete or non-differentiable functions, or black-box functions whose derivatives are not available, they cannot be expected to be of any value unless there is some correlation between goal function values at nearby candidate solutions—in other words, unless the goal function has a globally smooth continuous component more or less hidden by the jumps and bumps created by the discreteness constraints. Yet none of the popular meta-heuristics uses the know-how of continuous optimization when trying to exploit that continuous component. For example, if the problem is to find two integers that minimize the Q function above, known meta-heuristics (including genetic ones) will fail to notice the overall quadratic behavior of Q, and will essentially behave as a random local search—or worse. (Note that this remark refers to the global behavior of the goal function, not the local smoothness of a continuous goal function with many local minima. Such local smoothness is most effectively exploited by using continuous optimization methods inside the generator/mutator procedures, so that the meta-heuristic only sees a discrete search space consisting of the local minima.)