Applying Genetic Algorithms to Multi-objective Land-Use Planning


This thesis explores the potential of applying genetic algorithms (GAs) to multi-objective land-use planning. Rural land managers, in the UK and elsewhere in Europe, are faced with an increasingly complex decision-making environment where a varied mix of goals have to be achieved. To address this problem, GA-based land-use planning tools have been developed that interface with a decision support system. The thesis first evaluates the options for representing the land-use planning problem within a GA framework. Two genotype representations are proposed: a genotype mapping land-use directly to land parcels (Land-Block), and a representation making allocations indirectly via a greedy algorithm (P&P). The P&P representation requires novel breeding operators and post-evaluation processing, to identify and remove duplicate or defective genotypes. The performance of the two GAs' was evaluated for a single-objective land-use planning problem. Since both GAs found acceptable solutions, two multi-objective genetic algorithms (mGAs) were implemented based on the proposed representations. The goal of these mGAs was to search for a set of solutions defining the structure of the relationship between two or more objectives. To achieve this goal, modifications are required to the calculation of selection-fitness and the implementation of parent selection. The mGAs were tested for a problem with two conflicting objectives, with the Land-Block mGA found to have superior performance. Given the non-standard nature of the GAs, it was necessary to investigate their parameterisation, in particular the balance between the GA-operators over the course of the run. Two online-optimisation approaches were tested. The use of online-optimisation was successful in significantly improving the efficiency of the GAs. To investigate the usefulness of the GAs and mGAs for decision support, two further investigations were carried out. In the first the scalability of the two representations was tested for a single-objective problem. The P&P GA outperformed the Land-Block, since the complexity of the P&P genotype depends on the number of land-uses present in the optimal solution, not on the number of land parcels to be allocated a land use. Second, the sets of solutions found by the mGAs were compared with allocations collected, using soft-systems methods, from professional land-managers with a range of backgrounds. The mGAs found solutions as good as those proposed by the land managers. Additional factors that needed to be accounted for by the mGA were ide~tified, but the mGA solutions were seen by the land-managers as forming a useful basis for practical land-use planning.