Słownik AI
Kompletny słownik sztucznej inteligencji
Zero-Sum Game
Theoretical scenario where one agent's total gain exactly equals another's loss, fundamental in multi-agent adversarial learning to model strict competitions.
Minimax Algorithm
Decision-making algorithm that maximizes the minimum possible gain in adversarial situations, used to develop robust strategies against the opponent's worst actions.
Nash Equilibrium
Stable state where no agent can improve their strategy by unilaterally changing their behavior, crucial for analyzing equilibrium points in adversarial MARL.
Self-Play
Training methodology where an agent learns by competing against evolving copies of itself, eliminating the need for external data.
Robust Policy
Reinforcement learning policy that maintains high performance against adversarial perturbations or unexpected changes in the environment.
Worst-Case Optimization
Optimization paradigm aimed at maximizing performance in the most unfavorable scenarios, essential for developing agents resilient to adversarial attacks.
Adversarial Attack
Deliberate action by an agent aimed at degrading another agent's performance through environmental manipulation or injection of malicious perturbations.
Defense Strategy
Set of mechanisms and policies designed to detect, counter, and recover from adversarial attacks in multi-agent systems.
Adversarial Environment
Learning environment designed to actively present challenges and obstacles to agents, simulating hostile or unpredictable real-world conditions.
Policy Distillation
Knowledge transfer technique where a complex policy learned by an agent is compressed into a simpler and more efficient form, often used after adversarial training.
Adversarial Reinforcement Learning
Reinforcement learning paradigm explicitly integrating adversarial agents into the training process to improve robustness and generalization capabilities.
Multi-Agent Adversarial Bandit
Extension of the multi-armed bandit problem where multiple agents interact in an environment with rewards potentially manipulated by adversaries.
Adversarial Imitation Learning
Imitation learning approach using adversarial discriminators to evaluate and improve the quality of imitated behavior compared to experts.
Robustness Testing
Systematic evaluation of agent performance in extreme scenarios and coordinated attacks to measure their resilience and identify vulnerabilities.
Adversarial Perturbation
Subtle but intentional modification of observations or the environment designed to induce errors in a target agent's decision-making.
Strategic Uncertainty
Uncertainty about the future intentions and strategies of adversaries, requiring probabilistic and adaptive approaches in multi-agent decision-making.
Game-Theoretic MARL
Application of game theory to multi-agent reinforcement learning to analyze and optimize strategic behaviors in competitive contexts.