A hidden target needs to be found by a searcher in many real-life situations, some of which involve large costs and significant consequences with failure. Therefore, efficient search methods are paramount. In our search model, the target lies in one of several discrete locations according to some hiding distribution, and the searcher's goal is to discover the target in minimum expected time by making successive searches of individual locations.
In Part I of the thesis, the searcher knows the hiding distribution. Here, if there is only one way to search each location, the solution to the search problem, discovered in the 1960s, is simple; search next any location with a maximal probability per unit time of detecting the target. An equivalent solution is derived by viewing the search problem as a multi-armed bandit and following a Gittins index policy.
Motivated by modern search technology, we introduce two modes---fast and slow---to search each location. The fast mode takes less time, but the slow mode is more likely to find the target.
An optimal policy is difficult to obtain in general, because it requires an optimal sequence of search modes for each location, in addition to a set of sequence-dependent Gittins indices for choosing between locations. For each mode, we identify a sufficient condition for a location to use only that search mode in an optimal policy.
For locations meeting neither sufficient condition, an optimal choice of search mode is extremely complicated, depending both on the hiding distribution and the search parameters of the other locations.
We propose several heuristic policies motivated by our analysis, and demonstrate their near-optimal performance in an extensive numerical study.
In Part II of the thesis, the searcher has only one search mode per location, but does not know the hiding distribution, which is chosen by an intelligent hider who aims to maximise the expected time until the target is discovered. Such a search game, modelled via two-person, zero-sum game theory, is relevant if the target is a bomb, intruder, or, of increasing importance due to advances in technology, a computer hacker. By Part I, if the hiding distribution is known, an optimal counter strategy for the searcher is any corresponding Gittins index policy.
To develop an optimal search strategy in the search game, the searcher must account for the hider’s motivation to choose an optimal hiding distribution, and consider the set of corresponding Gittins index policies.
%It follows that an optimal search strategy in the search game must be some Gittins index policy if the hiding distribution is assumed to be chosen optimally by the hider.
However, the searcher must choose carefully from this set of Gittins index policies to ensure the same expected time to discover the target regardless of where it is hidden by the hider.
%It follows that an optimal search strategy in the search game must be a Gittins index policy applied to a hiding distribution which is optimal from the hider's perspective. However, to avoid giving the hider any advantage, the searcher must carefully choose such a Gittins index policy among the many available.
As a result, finding an optimal search strategy, or even proving one exists, is difficult.
We extend several results for special cases from the literature to the fully-general search game; in particular, we show an optimal search strategy exists and may take a simple form. Using a novel test, we investigate the frequency of the optimality of a particular hiding strategy that gives the searcher no preference over any location at the beginning of the search.