AI 용어집
인공지능 완전 사전
Pareto Optimum
Optimal solution in a multi-objective context that cannot be improved on any objective without degrading performance on at least one other objective.
Multi-Objective Reinforcement Learning
Extension of reinforcement learning where the agent simultaneously optimizes multiple often conflicting objectives with vector reward functions.
Vector Reward Function
Function that returns a vector of rewards instead of a scalar value, allowing simultaneous consideration of multiple performance criteria.
Objective Weighting
Scalarization technique where each objective receives a weight to combine multiple rewards into a single scalarized value to optimize.
Linear Scalarization
Method that transforms a multi-objective problem into a scalar problem through linear weighted combination of objectives to generate different Pareto-optimal solutions.
Pareto Elitism
Strategy that preserves non-dominated solutions between generations to ensure convergence to the Pareto front in evolutionary algorithms.
Expected Vector Return
Generalization of expected return in reinforcement learning, calculating the expectation of future cumulative reward vectors for each policy.
Pareto-optimal Policy
Action policy whose vector return belongs to the Pareto front, representing an optimal trade-off between different objectives.
Convergence Pareto
Propriété d'un algorithme garantissant que les solutions générées tendent asymptotiquement vers le véritable front de Pareto du problème.
Agrégation Tchebychev
Méthode de scalarisation utilisant la norme Tchebychev pour combiner les objectifs, capable de générer toutes les solutions Pareto-optimales convexes et non-convexes.