Forecasting Future World Events with Neural Networks
Introduces Autocast, a dataset for measuring the ability of NNs to forecast future world events. The results show significant room for future improvement.
This paper highlights the challenge of forecasting world events and introduces a dataset for measuring NNs' performance in this task. The use of Autocast could help improve policy and decision-making in various domains such as climate, geopolitical conflict, pandemics, and economic indicators. The paper also suggests that improved performance in NNs could contribute to large practical benefits.
Mastering the Game of Stratego with Model-Free Multiagent Reinforcement Learning
Introduces DeepNash, an autonomous agent capable of learning to play the imperfect information game Stratego from scratch, up to a human expert level.
This paper introduces an autonomous agent, DeepNash, that can play the game of Stratego up to a human expert level. The paper highlights the challenges of mastering Stratego and the complexity of decision-making under imperfect information. DeepNash uses a game-theoretic, model-free deep reinforcement learning method, without search, to learn to master Stratego via self-play. This could have implications for improving AI's ability to handle other games with similar challenges.