预测性智能合约

由A.I主导的奖励机制时代

David Dao
12 min readNov 29, 2017

Read the english version here

机器能引导我们做个好人吗?

趋利是人的天性。将人的固有的自利转化为有利于社会的行动是极具挑战性的,若能成功,这将颠覆整个时代。试想,如果为社会做贡献是对自己最有益的,它自然会成为人们的生活习惯。然而,如何设计出可持续并且诱人的双赢奖励机制则面临着诸多挑战:

  • 如何给社会事业评估其价值,并确保我们的奖励能够最大限度地促进这一事业?
  • 如何设计出足够灵活的奖励措施,以适应不断变化的情况?
  • 如何确定这些奖励措施是可靠的?

在这篇文章中,我们并非要领域专家亲自设计这些奖励措施,而是要介绍以人工智能代为设计奖励机制的技术。我们认为,人工智能驱动的方法不仅能使我们克服许多上述障碍,更能帮助我们设计更高效,可扩展和数据驱动的策略。

GainForest — 一个(非常)机智的打击亚马逊森林砍伐的合同

森林砍伐和森林退化所导致的碳排放量约占全球碳排放的17%,超过全球交通运输领域,仅次于能源领域

我们从一个激励型的例子开始:热带森林的砍伐和退化占全球温室气体(GHG)年排放量的17%左右,要避免危害全人类的气候变化,减少森林砍伐是必要的。在亚马逊地区,超过80%的森林砍伐的原因当地农民需要为种植和牧畜腾出空间。对护林员的小额经济激励措施可使森林砍伐减少一半。但是,认证谁来照顾亚马逊是一项非常耗费人力的任务。

在hack4climate中,我们团队意识到,可以利用经济利益激励人们成为护林员,因此,我们建立了GainForest。这是一个透明,可扩展的平台,允许任何人通过区块链式智能合约成为热带雨林地区的利益相关者。

护林员的安置制度

可从中获益者受利益所激励,自愿成为护林员。全球化的投资系统允许任何人成为护林员。

这个想法很简单:护林员负责确保亚马逊的某一片森林区域免受砍伐 。他们先选择一块区域并且投入一笔资金,在保护期结束后(1/3/6个月),如果他们选择的森林未受破坏,他们将获得最初的投注并且获得奖励。森林若遭破坏,他们的投注将用来支付给以后的护林员。奖励根据保护的难易水平,该区域中有多少利益相关者,以及其他因素来计算。有经济利益的人会受激励来看护他们在本地或全球投资的地区。当地农民可以简单地投资自身所在的区域,在当地亲自看护森林,而全球企业家或社区可以投资,并帮助制定政策,并支付当地的护林员,以打击当地的森林砍伐行为。

此外,捐助者可以通过金钱捐赠系统将资金重新分配给护林员。由于透明的智能合约再分配系统,他们可以即时地看到他们的援助在何时何地取得何等成效。

基于区块链的智能合约,交易将是安全透明的

人工智能机器激励系统概况

智能合同是一种计算机协议,在签署协议后,如果所有条件都得到满足,它保证会被精确地遵守与履行。重要的是,智能合约是自动生效和自动执行的。它们被部署在一个分散的区块链上,消除了对可信赖的第三方(人)的需求。所有交易也透明地记录在区块链中,使捐赠者能够掌握捐赠对环境的影响。

更进一步,通过将人类排除在重新分配的循环过程外,我们可以将其扩展成为一个更加透明的数据驱动系统。

利用机器智能进行数据驱动的投注分配

如前所述,护林员的认证及其在亚马逊的影响是一项非常耗费人力的任务。我们提出了一个基于机器学习的oracle,可以自动评估Amazon的每块区域单元。例如,遥感算法可以轻易从卫星图像(如GLAD,DETER,SAD或FORMA)中检测森林砍伐。

GLAD是一个每周砍伐森林警报系统。它从卫星图像中提取信息,并能够在30m分辨率内检测树木覆盖的减少量(粉色)。这个数字显示了从2015年1月1日到2017年11月1日的树木减少情况。

除此之外,我们还可以进一步利用机器智能。投标制度有以下弱点:一个狡猾的投资者可以把所有的钱投入到亚马逊的一个偏远地区,因为他知道森林被砍伐的可能性非常低。因此,我们需要找到一种将利率与森林砍伐的实际风险联系起来的方法。总之:我们需要预测驱动的奖励。在黑客马拉松(Hackathon)期间,我们从GLAD收集了可取得的图像和森林砍伐数据,并训练了一个卷积神经网络来预测未来森林被砍伐的几率。令我们惊讶的是(考虑到24小时的时间限制),我们能够得到合理的预期结果。我们认为如果有更多的数据和更长的时间,该系统可以计算出更精确的结果。

尽管深度神经网络并不能精准预测森林砍伐,但我们可以看到它学习到了有效的模式。也许它只是预测了一个更长远的未来? (左:地表实况,右:森林砍伐的预测)

我们制作了:GainForest.org,一个区块链激励预测智能合同。我们赢得了hack4climate的一等奖,并在2017年联合国气候变化大会(COP23)上发表。欢迎访问我们的网站,并查看我们的代码。

Predicting more applications for smart contracts

sustainable development

Our satellite imagery approach can be flexibly applied to more environmental scenarios such as river pollution, oil palm cultivation and marine conservation.

traffic department

Traffic congestion and rush hour forecasting are very common tasks in modern smart cities. If they can be combined with smart contracts, they can dynamically adjust the fares of the public transport system to motivate people to use buses rather than private cars.

Criminal investigation

An exciting application may be to identify potential incentives to prevent crime from rising. Based on historical data on policies and their impact on certain urban areas, artificial intelligence can be used to design new policies and incentives to reduce the crime rate.

Microfinance

It is more difficult to verify the impact of microfinance. By linking existing infrastructure and economic data with microfinance incentives in certain areas, it is possible to identify regions with the highest growth potential and allow these regions to engage in low-risk loans.

What is the next step? Learn Strategic Inspiration from Data

Blockchain and artificial intelligence are not related. However, from GainForest's example, we see that artificial intelligence can help us to automatically assess risk through forecasting, so that we have dynamic smart contracts. On the other hand, smart contracts help us define a system of self-enforcement and constraint, and it is also a perfect attempt to use neural networks for prediction. We believe that with the advancement of artificial intelligence and blockchain technology, predictive smart contracts will play a huge role in the future of blockchain trading systems. In the case of GainForest, to a certain extent, we can even use limited data to predict deforestation patterns in different parts of the world. That's because artificial intelligence algorithms like neural networks can transfer learned knowledge from one data set to another. Meta-learning thus allows us to train in the Amazon rainforest (a region with large data volumes) and then apply the prediction to areas where other data in the world are scarce.

Even more interesting is that our neural network is sending out self-defeated prophecies. When our learning algorithm predicts deforestation to a certain degree, it will prevent predictions because it is also included in the prediction system. So why do we not only optimize the accuracy Basically networks can motivate rangers and harvesters to try to maximize benefits. Basically, neural networks want to expand the protected areas of tropical Rain forests, while the harvesters try to minimize it. With this goal of game theory, maybe neural networks can learn predictive Incentive strategies and have positive guidance for human behavior. At the confluence of deep learning,Blockchain technology and game theory, there must be an exciting area of ​​research that is waiting to be explored!

Acknowledgements

Many thanks to Shan for the chinese translation and Yunyan and Shaoduo for proofreading it!

Links

Research Group

Pitch Slides

Code

--

--

David Dao

PhD student in AI and Data Systems for Sustainable Development 🌱🛰️🌍 | Founder GainForest.app | Past: Stanford, Berkeley, MIT https://daviddao.org