本文探讨了机器学习、强化学习和决策智能领域中不同类型的决策，包括AML分类、BML回归、决策智能（前向模型和优化）、强化学习等。作者指出，决策智能领域存在着巨大的未满足需求，即在没有完整数据的情况下，需要做出具有重大影响的决策。决策智能旨在将信息转化为更好的行动，并回答“如果今天做出这个决策，导致这个行动，明天会有什么结果？”的问题。文章最后引发了关于决策智能是否应该扩展到包括AML分类和BML回归的讨论。

🤔**AML分类**：主要用于回答“这是什么图片？”、“这个人患了什么病？”等问题，目标是通过数据进行分类，例如识别猫或诊断疾病，其成功标准是真阳性率和真阴性率，通常使用监督学习方法。

📊**BML回归**：主要用于预测未来，例如预测下个月的疫情发生率或股票价格，目标是通过数据进行预测，其成功标准是均方误差和R平方，通常使用监督学习方法。

💡**决策智能（前向模型）**：关注的是行动与结果之间的映射关系，即“如果我采取这个行动，在这个情境下，结果会是什么？”，需要结合人类经验、机器学习、经济学等多种信息来源构建因果模型，其成功标准是正确映射行动到结果，通常使用复杂系统模拟方法。

🎯**决策智能（优化）**：在给定的行动集合中，寻找最优的行动组合以实现目标，例如找到最优的决策来实现多目标的最优化，其成功标准是达到多目标的最优结果，通常使用复杂系统模拟和优化方法。

🤖**强化学习**：旨在创建策略，即在每个可能的状态下，确定最佳的下一步行动，其成功标准是最大化目标函数的价值，通常使用强化学习模拟方法。

In artificial intelligence, machine learning, decision intelligence, statistics, and science, we use the word “decision” to mean a lot of things. Let’s tease out some distinctions:

Decision Type	Name	Question answered	Primary information Source	Typical success criterion	Typical method
A	ML classification	“Decisions That”: “What is this picture?” “What disease does this person have?” “Is this a cat?”	Data	True positive, true negative	Supervised learning
B	ML regression	“Decision about a prediction”: “What will be the Covid-19 incidence next month?” “What will be this security’s price next month?”	Data	Mean squared error, R^2	Supervised learning
C	Decision intelligence (forward model)	Decision to take an action, action-to-outcome mapping: “If I take this action, in this context, what will be the outcome?”	Humans (causal model), ML, economics, complex systems models, much more (causal model links)	Correct mapping of actions to outcomes	Complex systems simulation
D	Decision intelligence (optimization)	“Given my set of possible actions, what is the best set of actions to take to meet my goals”	(same as above)	Best set of decisions to reach multi-objective outcomes	Complex systems simulation, optimization
E	Reinforcement learning	Policy creation: “For each state that I can be in, what is the best next action?” (policy)	Data and simulation	Best set of policies to maximize value of objective function	Reinforcement learning simulation

Cassie Kozyrkov and I have realized that there’s unmet need to fill the space marked “DI”, above, where:

We don’t necessarily have data for the entire decision, yet…people need to make this decision today, without waiting to have time to gather the data, and…the decision has big impact, and…we want to make better decisions, and we have some, if not all, data and human expertise available to us yet it’s not being well utilized.

We, along with a few thousand others, have realized that this is an important, yet massively under-treated, corner of the decision problem formulation space.

You can learn about it in this course: Getting Started with Decision Intelligence.

“Decision intelligence is the discipline of turning information into better actions at any scale.” —Cassie Kozyrkov, Head of Decision Intelligence, Google

“Decision intelligence answers the question, ‘If I make this decision today, which leads to this action, what will be the outcome tomorrow?’—Lorien Pratt, Chief Scientist, Quantellia

Do you agree? Lately, I’ve heard a case that “Decision Intelligence” should be expanded to include both A and B above.  What do you think?

Thanks to @thenatlog, @neuralnets4life, and @twimlai for inspiring this post