Multi-Criteria Decision Making (MCDM)

TOPSIS is a compensatory, distance-based MCDM method that selects the alternative closest to an ideal solution and farthest from a negative-ideal solution. The algorithm proceeds in distinct steps:

1. Construct a normalized decision matrix.
2. Create a weighted normalized matrix using criteria weights.
3. Determine the ideal (best performance) and negative-ideal (worst performance) points.
4. Calculate the Euclidean distance of each alternative to these two reference points.
5. Rank alternatives by their relative closeness coefficient.

TOPSIS is computationally straightforward and provides a clear geometric interpretation of the 'best compromise' solution.

PROMETHEE is an outranking method that builds a preference structure by comparing alternatives pairwise for each criterion. It uses preference functions (e.g., linear, Gaussian) to translate the difference in performance between two alternatives into a degree of preference (from 0 to 1). Key outputs include:

• Positive and negative outranking flows, representing how much an alternative outranks, or is outranked by, all others.
• A net flow used for a complete ranking (PROMETHEE II).
• A partial preorder (PROMETHEE I) that can reveal incomparabilities.

PROMETHEE is valued for its flexibility in modeling nuanced decision-maker preferences through its choice of preference functions.

The ELECTRE family of methods employs an outranking relation to model the decision-maker's possible hesitation between alternatives. It is a non-compensatory approach, meaning a poor score on one criterion cannot be fully offset by excellent scores on others. The core process involves:

• Defining concordance and discordance indices for each pair of alternatives.
• Concordance measures the coalition of criteria supporting the assertion that one alternative is at least as good as another.
• Discordance measures the strength of criteria that strongly oppose this assertion.
• Building outranking relations based on thresholds, leading to a partial ranking that may include incomparable alternatives, reflecting real-world decision complexity.

This category encompasses iterative, human-in-the-loop approaches. Goal Programming seeks to minimize deviations from pre-defined target levels for each objective. Interactive Methods, such as the Step Method (STEM), involve a cyclical process:

1. The algorithm presents a candidate solution from the Pareto front.
2. The decision-maker provides feedback, such as relaxing a goal for one objective to gain improvement in another.
3. The algorithm uses this preference articulation to refine the search space.

These methods are essential when decision-maker preferences are not fully known a priori, allowing for progressive learning and refinement of the most desirable trade-offs.

Methods that transform a multi-objective problem into a single-objective problem, allowing the use of traditional optimization algorithms. This is a primary method for preference articulation.

• Weighted Sum Method: Aggregates objectives into a single score using a weighted linear combination. Critically, it cannot find solutions on non-convex parts of the Pareto front.
• Epsilon-Constraint Method: Optimizes one primary objective while converting others into constraints with allowable bounds (epsilon values).
• Goal Programming: Minimizes the deviation from a set of predefined target levels or goals for each objective.

A class of population-based metaheuristics designed to approximate the full Pareto front in a single run. They are particularly effective for complex, non-linear, or discontinuous problems.

• NSGA-II (Non-dominated Sorting Genetic Algorithm II): A seminal algorithm that uses non-dominated sorting for selection and crowding distance to maintain diversity along the front.
• MOEA/D (Multi-Objective EA Based on Decomposition): Decomposes the problem into many single-objective subproblems using scalarization and solves them cooperatively.
• Archive: A common component that stores the best non-dominated solutions found during the evolutionary search.

A structured, pairwise comparison technique for organizing and analyzing complex MCDM problems. It breaks a problem into a hierarchy of criteria, sub-criteria, and alternatives, then uses human judgments to assign weights and priorities.

• Pairwise Comparisons: Decision-makers evaluate the relative importance of elements using a standardized scale (e.g., 1-9).
• Consistency Ratio: A mathematical check to ensure the judgments are logically consistent.
• Synthesis: Calculates global priorities for each alternative by aggregating weights throughout the hierarchy.

A subfield of reinforcement learning where the agent receives a vector-valued reward signal representing multiple objectives. The agent must learn policies that effectively balance these competing goals over time.

• Single-Policy vs. Multi-Policy: Approaches can seek a single policy that satisfies a specific scalarization (e.g., weighted rewards) or a set of policies covering the Pareto front.
• Challenge: The credit assignment problem becomes more complex, as the agent must understand the long-term trade-offs of actions across multiple reward dimensions.
• Application: Ideal for autonomous systems (e.g., robots, trading agents) that must balance efficiency, safety, and cost in real-time.