Counterfactual state explanations for reinforcement learning agents via generative deep learning

• Deep learning models can generate counterfactual states for game-playing agents. • Generative counterfactual states are sufficiently realistic according to humans. • Counterfactual states can be used for identifying a flawed agent. • Generated counterfactual states outperform nearest neighbors in...
Ausführliche Beschreibung

• Deep learning models can generate counterfactual states for game-playing agents. • Generative counterfactual states are sufficiently realistic according to humans. • Counterfactual states can be used for identifying a flawed agent. • Generated counterfactual states outperform nearest neighbors in flaw detection. Ausführliche Beschreibung