Global S&T Development Trend Analysis Platform of Resources and Environment
| DOI | 10.1126/science.abi7668 |
| Machine-generated theories of human decision-making | |
| Sudeep Bhatia; Lisheng He | |
| 2021-06-11 | |
| 发表期刊 | Science
 |
| 出版年 | 2021 |
| 英文摘要 | Imagine a choice between two gambles: getting $100 with a probability of 20% or getting $50 with a probability of 80%. In 1979, Kahneman and Tversky published prospect theory ([ 1 ][1]), a mathematically specified descriptive theory of how people make risky choices such as these. They explained numerous documented violations of expected utility theory, the dominant theory at the time, by using nonlinear psychophysical functions for perceiving underlying probabilities and evaluating resulting payoffs. Prospect theory revolutionized the study of choice behavior, showing that researchers could build formal models of decision-making based on realistic psychological principles ([ 2 ][2]). But in the ensuing decades, as dozens of competing theories have been proposed ([ 3 ][3]), there has been theoretical fragmentation, redundancy, and stagnation. There is little consensus on the best decision theory or model. On page 1209 of this issue, Peterson et al. ([ 4 ][4]) demonstrate the power of a more recent approach: Instead of relying on the intuitions and (potentially limited) intellect of human researchers, the task of theory generation can be outsourced to powerful machine-learning algorithms.
The popularity of prospect theory led to new research programs in psychology, economics, business, and neuroscience, as well as to the development of descriptive models for domains like intertemporal, social, strategic, and consumer choice ([ 5 ][5]–[ 7 ][6]). Prospect theory also helped practitioners and policy-makers derive practical insights on how to improve individual and organizational decision-making ([ 8 ][7]). But the prospect theory approach to modeling choice behavior is not without drawbacks. Researchers who propose new theories usually make complicated assumptions about processes such as perception, attention, memory, and emotion, as well as sources of noise and error in choice. The theories themselves are tested only on small datasets of choices and are seldom compared against the large set of preexisting models. This is unavoidable, given the long interdisciplinary history of decision research and the complexity of risky choice: It is fairly easy for decision scientists to intuit a psychological explanation for an expected utility violation, but even the most talented (human) theorist will have difficulty deriving predictions that distinguish their account from dozens of preexisting explanations. Newer theories are often similar to previously published models, and many theories closely mimic each other's predictions on benchmark datasets ([ 3 ][3]). Even though the rate of theory production is accelerating, there has been little gain in predictive accuracy on these datasets over the past 20 years ([ 9 ][8]).
In response to these trends, some researchers have begun to emphasize out-of-sample predictions and comprehensive model comparisons on large datasets of human decisions ([ 10 ][9], [ 11 ][10]). In doing so, they have brought yet another discipline, machine learning, into this already highly interdisciplinary mix. Old theories need not be discarded; rather, they can be implemented as inductive biases or constraints in these models, increasing algorithms' learning efficiency and boosting model performance.
Peterson et al. showcase the true power of this approach. They begin by collecting new experimental data on risky decisions for more than 10,000 distinct choice problems involving gambles with probabilistic monetary payoffs, exceeding the size of prior datasets by an order of magnitude. These choice problems and the decisions that humans ultimately made in these problems are then used to train deep neural networks, a class of machine-learning models that can flexibly extract nonlinear functions for describing data. Peterson et al. find that such networks are able to mimic human decisions with a very high accuracy rate, substantially outperforming existing (human-generated) risky choice models; their model achieved roughly half the mean squared error demonstrated by prior approaches. In learning to mimic human decisions, the networks also discover many of the psychological properties of established behavioral theories, such as the psychophysical functions used by prospect theory. The flexibility of deep networks allows them to find better mathematical implementations for these properties and learn other properties necessary for describing data that have not been previously identified by human researchers.
The predictive gains of Peterson et al. relative to existing human-generated theories are impressive enough; however, the authors also analyze the modeling assumptions that lead to good performance. To do this, they implement various constraints on their networks, each of which limits the ways in which the networks manipulate the available gambles to make choices. For example, one constrained network allows for a payoff to be transformed nonlinearly and be multiplied against its (nontransformed) probability, resembling the assumptions of expected utility theory. This network performs poorly relative to a more complex network that allows for both probabilities and payoffs to be transformed, as in prospect theory. This network is, in turn, outperformed by other networks that also allow for the payoffs and probabilities of different gambles to interact and influence each other's transformations.
The architecture of the winning model in Peterson et al. 's analysis places it among a class of context-dependent decision theories that propose that the utility or disutility obtained from a single gamble (e.g., $100 with 20% probability) is not determined in isolation but depends on the other gambles available in the choice problem (e.g., $50 with 80% probability). Decision-makers do not attach a fixed utility to any given gamble; rather, utility is based on competing gambles and can vary from choice problem to choice problem.
Peterson et al. also show that high accuracy rates are possible from a model composed of an expected utility theory component and a prospect theory component, with each component being selectively applied based on the set of available gambles. This analysis illuminates how decision context guides and constrains evaluation mechanisms and shows how interpretable theoretical insights can be obtained from the behavior of deep neural networks.
Future work will undoubtedly extend Peterson et al. 's approach to other behavioral domains [e.g., risk perception, moral judgment, and strategic choice ([ 12 ][11]–[ 14 ][12])]. These models will also be used to discover new choice problems to test the boundaries of existing theories. Advances in explainable artificial intelligence will also allow researchers to better understand the behavior of deep networks in terms of established theoretical principles. Ultimately, the increased availability of large datasets and improvements in computing power will make machine learning an indispensable component of the decision scientist's toolbox, revitalizing (and perhaps, once again, revolutionizing) theoretical research on human choice behavior.
1. [↵][13]1. D. Kahneman,
2. A. Tversky
, Econometrica 47, 263 (1979).
[OpenUrl][14][CrossRef][15][PubMed][16][Web of Science][17]
2. [↵][18]1. W. Edwards
, Annu. Rev. Psychol. 12, 473 (1961).
[OpenUrl][19][CrossRef][20][PubMed][21][Web of Science][22]
3. [↵][23]1. L. He,
2. W. J. Zhao,
3. S. Bhatia
, Psychol. Rev. 10.1037/rev0000231 (2021).
4. [↵][24]1. J. C. Peterson,
2. D. D. Bourgin,
3. M. Agrawal,
4. D. Reichman,
5. T. L. Griffiths
, Science 372, 1209 (2021).
[OpenUrl][25][Abstract/FREE Full Text][26]
5. [↵][27]1. J. W. Payne,
2. J. R. Bettman,
3. E. J. Johnson
, The Adaptive Decision Maker (Cambridge Univ. Press, 1993).
6. 1. C. F. Camerer,
2. G. Loewenstein,
3. M. Rabin
, Eds., Advances in Behavioral Economics (Princeton Univ. Press, 2004).
7. [↵][28]1. P. W. Glimcher,
2. E. Fehr,
3. R. A. Poldrack
, Eds., Neuroeconomics: Decision Making and the Brain (Academic Press, 2013).
8. [↵][29]1. R. H. Thaler,
2. C. R. Sunstein
, Nudge: Improving Decisions About Health, Wealth, and Happiness (Penguin, 2009).
9. [↵][30]1. L. He,
2. P. P. Pantelis,
3. S. Bhatia
, “The wisdom of model crowds” (2021); |
| 领域 | 气候变化 ; 资源环境 |
| URL | 查看原文 |
| 引用统计 | |
| 文献类型 | 期刊论文 |
| 条目标识符 | http://119.78.100.173/C666/handle/2XK7JSWQ/329901 |
| 专题 | 气候变化 资源环境科学 |
| 推荐引用方式 GB/T 7714 | Sudeep Bhatia,Lisheng He. Machine-generated theories of human decision-making[J]. Science,2021. |
| APA | Sudeep Bhatia,&Lisheng He.(2021).Machine-generated theories of human decision-making.Science. |
| MLA | Sudeep Bhatia,et al."Machine-generated theories of human decision-making".Science (2021). |
| 条目包含的文件 | 条目无相关文件。 | |||||
| 个性服务 |
| 推荐该条目 |
| 保存到收藏夹 |
| 查看访问统计 |
| 导出为Endnote文件 |
| 谷歌学术 |
| 谷歌学术中相似的文章 |
| [Sudeep Bhatia]的文章 |
| [Lisheng He]的文章 |
| 百度学术 |
| 百度学术中相似的文章 |
| [Sudeep Bhatia]的文章 |
| [Lisheng He]的文章 |
| 必应学术 |
| 必应学术中相似的文章 |
| [Sudeep Bhatia]的文章 |
| [Lisheng He]的文章 |
| 相关权益政策 |
| 暂无数据 |
| 收藏/分享 |
除非特别说明，本系统中所有内容都受版权保护，并保留所有权利。
修改评论