OPTAIR:     Multi-criteria optimization for air quality
management and emission control

4. Technical challenges

1. Computational performance and combinatorial explosion. The very large number of possible emission control technology/source combinations, even when clustering sources by category that needs to be treated in parallel, together with the considerable compute times for the air quality models themselves makes any exhaustive search impossible. Solutions to be tested include:
   • Use of multiple models: perform first layer of screening runs with a simpler model (e.g., AERMOD, PBM), then analyse only the most promising candidates with a full resolution 3D dynamic model;
   • Use of selected periods: concentrate on “worst case” meteorological conditions, optimize for those, then check feasibility only for a few non-dominated alternatives with the full temporal coverage of one year at hourly resolution designed to provide key indicators like “number of hourly violations”.
   • Explore grid computation possibilities beyond parallel cluster computing, involving computational resources at selected EUREKA project partner institutions.
   • Develop adaptive heuristics to increase the efficiency of the search: one possibility beyond classical genetic algorithms is to use machine learning concepts (e.g., ID3/ID5R) that can pre-select most promising instruments.
2. Convergence: a major problem and thus risk of the proposed optimization approach is that there is no guarantee of convergence or global optimality, which is always probabilistic: however, the satisficing paradigm does address this dilemma (Byron, 1998; see e.g., Conley 1980 as one of the first books on numerical optimization techniques). The most promising approach to address this problem is:
   • Increased performance to increase the set of feasible solutions, increasing the probability to approach a global optimum;
   • Development of efficient tie-breaking mechanisms to avoid oscillation around local maxima in any local random-walk hill climbing part of the strategy.
3. Uncertainty and the stochastic nature of the boundary conditions: Ambient air quality results from the interplay of emissions, and the meteorological boundary conditions. Both are known or predicted with a measurable level uncertainty. Uncertainty may be rather low like in the case of a major thermal power stations, or very high, in particular for distributed behavioral components like traffic.

   Meteorological forecasting at the local scale is notoriously uncertain, in particular, when we consider long-term climate change as one of the driving forces, where any forecast can only be statistical in nature. To implement solutions as the basis for policy making that require medium- to long-term investments requires a high level of reliability and robustness. WP 12 will analyze issues of uncertainty and robustness, including sensitivity of solutions to climate change scenarios downscaled from IPCC GCM/SRES scenario (A2 and B2 will be used to drive the prognostic MM5 model): in short, solutions that are pareto optimal over a wide range boundary conditions can be considered as robust. Uncertainty or at least great temporal variability in the dynamic boundary conditions (meteorology) make it even more difficult to find a reliably optimal strategy for emission reduction and pollution control, as we expect such solutions to be optimal over the long run, but good enough (feasible) at any time. Depending on the dynamics of the system, only an adaptive solution may be able to meet this requirement: a classical even if somewhat helpless example is to stop urban traffic during smog or ozone episodes. A specific work package (WP 12) will analyse this issue in the second and third project year.

4. Valuation and end user (DM) involvement: multi-criteria optimization necessarily involves a subjective element at the very least in the trade off between criteria, but already in the choice of criteria, and the a priori constraints. The corresponding value judgements (even if based on legal and regulatory constraints or targets set by law, which is still at best inter-subjective) are in the realm of socio-political if not group-psychological process rather than natural sciences, engineering, or mathematics. The seamless yet credible and scientifically rigorous integration of natural sciences and socio-economic considerations in environmental management need:
   • An appropriate language or ontology and representation formats to capture and process perceptions, believes and aspiration in an unbiased, equitable and open, democratic process;
   • An interface that facilitates direct involvement of representative actors, stakeholders, and any legitimate participant in the decision making process.
   The key issue is a good balance and yet clean separation between neo-positivism and social constructivism, keeping each cleanly on their respective side of the decision making process: however, a practical DSS certainly needs both components.