This thesis investigates the extent to which multi agent based social simulations can validate the results or investigate the consistency of scenarios. The representation and analysis of future scenarios used by the Environment Agency (the Agency) provides an example of how Multi-Agent Systems can be used successfully where other techniques cannot.
Strictly qualitative or quantitative methods have been used to tackle the social aspects of water conservation (defined as “any beneficial reduction in water use or water losses”), but with little success. When existing quantitative (statistical) models were checked against recently (more recent than those that were used to generate the model) reveal that those that did best against the old data do worst with the new. On the other hand, it is extremely difficult to validate qualitative models against other results (such as those that can be expected from a formal representation).

It is argued that statistic methods failed because the standard assumption of normality does not hold in the data (as an analysis shows). Moreover, the properties of the system analysed and the fact that it generates power law distributed data suggests a conclusion that self-organised criticality is present.

A new model of domestic water demand is presented. Created in SDML (Strictly Declarative Modelling Language), the model attempts to represent the different scenarios used by the Agency in their water demand forecasts. Households are represented using a Multi Agent System (MAS): their characteristics and the way they interact with each other depend on the governance system and the social values, which are the drivers for each scenario. Agents are used to represent households, whose decisions can be analysed and the reasons for these decisions can be found via the database kept throughout the simulation. The four scenarios are differentiated in the model according to the governance system and the values in the society. Drivers for the model devised by the Environment Agency are: the water policy (represented via the policy agent and the regulations regarding appliances), the technology (represented via an agent’s personal and subjective beliefs), the behaviour (expressed as rules followed by households), and the economics (represented in the rate of replacement of appliances).

The model is used to investigate the extent to which the characteristics of the Agency’s four scenarios can be reproduced by the model. While the increase in water demand observed in scenarios A (Provincial Enterprise) and B (World Markets) is not present, scenarios C (Global Sustainability) and D (Local Stewardship) seem to better match the expected values. The fact that these scenarios take innovation into account tends to demonstrate a successful implementation of innovators in the model. In all scenario simulations as well as in the analysis of real data, the kurtosis (a measure of the peakedness for a distribution) does not correspond to normally distributed data. This explains that techniques based on this standard statistical assumption cannot be used, and demonstrates that this modelling conserve this property (or lack of thereof).

Sensitivity analysis leads to the conclusion that overall patterns observed did not seem significantly influenced by the grid structure of the model. The density of agents is more significant, as it requires a minimum density for the processes to take place, but further variations do not lead to changes in the underlying distribution of the water use data generated. On the other hand, the visibility for every agent is an important parameter. As the vision range increases, so does the information available to the agent. With more information comes an increase choice of actions, and as a consequence, potentially different outputs (namely increased vision range seems to result in a lower water demand).

Finally, whilst a direct comparison with the data and assumptions provided in the Agency’s “Water Resources for the Future – A Scenario Approach to Water Demand Forecasting” was not possible, this research shows that it is possible to obtain relatively similar results to the Agency’s in a coherent and consistent way while generating data with characteristics that are closer to the observed data than the existing statistical models. This also provides a better understanding of the target processes, as well as more possibilities for validation.