Smoke and Salt
Smoke and Salt is a real-time simulation that realizes artificial intelligence training, based on deep reinforcement learning algorithms. With the laws of physics, this project constructs a simulated environment in which a population of artificial life is “fed”. Each intelligent agent has a limited vision range; they observe the food hidden in the “hydrothermal vent” and act according to the specific location of the food. Similar to the survival mechanism in the real world, Smoke and Salt’s algorithm also sets a reward-and-punishment mechanism: points are added when the agent eats food; points are deducted if it has not eaten any food as time passes, and when the points drop to zero, the agent dies. The evolution of natural life has gone through billions of years, but fortunately, the learning process of artificial life is not as long. After hundreds of thousands of deaths with trial and error, the intelligent agent gradually becomes able to adapt to the complex environment. During testing, any small changes in the environment – such as the level of rewards and punishments, and the amount and location of food – may lead to completely different behaviors and fates of the agents.
《烟与盐》是一个基于深度强化学习算法来实现人工智能训练的实时 模拟。项目构建了一个有物理模拟的环境，在其中＂喂养＂了一群人工 生命。每个智能体拥有有限的视力范围，观察隐藏在＂海底热泉＂的食 物并依据食物的具体位置发起行动。与现实世界的生存机制相似，《烟 与盐》的算法也制定了一套赏罚机制：智能体吃到食物会加分；随时间 流逝，一直没有吃到食物会减分，当分数降至零时智能体便死亡。 自然生命的演化经历了数十亿年，幸运的是，人工生命的学习历程并 不如此漫长，在数十万次的死亡和试错后，智能体逐渐得以适应复杂环境。 调试时，赏罚的高低，食物的数量和位置等，任何一个环境中的微小改 变都有可能导致智能体截然不同的行为和命运