Give Me a One-Handed Economist: The Ongoing Struggle Between Social and Exact Sciences In Economy

As the USA president, Harry Truman met with many economists over the years [1]. Allegedly, at some point, Truman had enough and asked his office manager: "Give me a one-handed Economist! All my economists say 'on one hand...', then 'but on the other hand...'". In this immortal quote Truman was able to comically describe one of the major difficulties in economics: the dissonance between social sciences and the exact sciences. On one hand, social sciences are mostly empirical-based with minor to no theories. On the other hand, exact sciences rooted in theories increasingly being built from simple rules. As such, the economical science like no other science has to bridge between complex equations and computational models, and the massiness of working with systems involving people.

One direction economics took in the last few decades is behavioral economy [2]. This approach was established by the Nobel laureate Richard Thaler [3]. Behavioral economics combines elements of economics and psychology to understand how and why people behave the way they do in the real world. It differs from neoclassical economics, which assumes that most people have well-defined preferences and make well-informed, self-interested decisions based on those preferences. In plain English, behavioral economics examines the differences between what people "should" do and what they actually do.

While this direction seems promising and indeed yields a few revolutionary results, it is not without flaws. For example, a detailed review of papers in behavioral economy from 2018 shows almost no mathematical models to explain economical processes or phenomena [4]. The field mostly leans on social experiments and phycological theories. These have received a lot of criticism over the years due to the inability of the scientific community to reproduce results. This means behavioral economics gives interesting and useful information but does not provide a comprehensive understanding of economic processes. Therefore, it limits our ability to generalize known outcomes to new population groups and unfamiliar situations.

The challenges with this approach come from the inherent property of economic research. Economic methodology is limited by counterfactual data, simplistic behavioral models, and offers limited opportunities to experiment with policies. Thus, one can formally describe these challenges as follows: a central planner (for example, government) aims to find a policy under which the rational behavior of the affected economic agents (for example, people and businesses) yields desired social and financial outcomes. The classical and more modern economical approaches are usually limited by analytical traceability. The usage of assumptions makes these models easier (some would say feasible) to understand and analyze. However, these assumptions result in models that fail to capture the complexity of the real world.

Nonetheless, everything is not lost and economists can bridge the gap between social and exact sciences using computer science methods [5]. It is common to see how computer science bridges several sciences with mathematical models through applied algorithms [6], computational models [7], and in silico experiments [8]. A few key examples are: 1) Microbiology - calculating average nucleotide identity [9]; 2) DNA memory storage - modeling in vivo molecular memory systems that record and store information within the DNA of living cells [10]; and 3) Weather prediction - a machine learning approach [11].

To be specific, machine learning and computational techniques for automated mechanism design hold promise for overcoming the existing challenges. While promising, there has not been a general computational approach for policy design. Several attempts from the last years include genetic programming [6], Reinforcement learning [12], and agent-based simulations [13]. All these methods deal with a more abstract requirement of a realistic economic setup. There is a need for solving a highly non-stationary, sequential decision-making problem where all the agents are learning and adapting all the time. Moreover, different agents in the system have different goals. For example, consumers would like to gain as much as they can from their money while a government would like to obtain more taxes. This additinal layer makes these systems even more complex.

To conclude, recent studies suggest an exciting research agenda: using AI to enable a new approach to economic design. A Machine learning-based agents with a social agent-based simulation approach provide a robust framework for multiple economic processes. It can be used to study different policy goals and constraints, and, as AI-driven simulations grow in sophistication, may help to address modern economic questions. In particular, AI-driven simulations enable economic policies to be tested in more realistic environments than those available to analytical methods. However, before you all run to pop up a bottle of champagne, the known results are the first step and are not ready to be implemented as a real-world policy. Future research should scale up AI-driven simulations and calibrate them to real-world data. Specifically, the author of this blog post interesting in using this approach in a black market context.

References

1. The White House Offical Website
2. Pesendorfer, W. Behavioral Economics Comes of Age: A Review Essay on Advances in Behavioral Economics. Journal of Economic Literature. 2020; 44:3, 712-721.
3. Thaler, R. Toward a positive theory of consumer choice. Journal of economic behavior and organization. 1980; 1:1, 39-60.
4. Huang, Z., Gao, F. Literature Review of the Causes of the Vulnerability of Poverty in Traditional Economics and Behavioral Economics. 2020 International Symposium on Frontiers of Economics and Management Science. 2020.
5. Yec, C-H., Chen, S-H. On the Coordination and Adaptability of the Large Economy: An Application of Genetic Programming to the Cobweb Model. IFAC Proceedings Volumes. 1996; 29:1, 5659-5664.
6. Bhardwaj, R., Nambiar, A. R., Dutta, D. A Study of Machine Learning in Healthcare. 2017 IEEE 41st Annual Computer Software and Applications Conference. 2017; 236-241.
7. Prusinkiewicz, P., Runions, A. Computational models of plant development and form. New Phytologist, 2012; 193:1, 549-569.
8. Sotomayor, M., Schulten, K. Single-Molecule Experiments in Vitro and in Silico. Science. 2007; 316:5828, 1144-1148.
9. Lee, I., Kim, Y. O., Park, S-C., Chun, J. OrthoANI: An improved algorithm and software for calculating average nucleotide identity. International Journal of Systematic and Evolutionary Microbiology. 2016; 66:2.
10. Ceze, L., Nivala, J., Strauss, K. Molecular digital data storage using DNA. Nature Reviews Genetics. 2019; 20, 456-466.
11. Haupt, S. E., Cowie, J., Linden, S., McCandiess, T., Kosovic, B., Alessandrini, S. Machine Learning for Applied Weather Prediction. 2018 IEEE 14th International Conference on e-Science. 2018.
12. Mosavi, A., Faghan, Y., Ghamisi, P., Duan, P., Ardabili, S. F., Salwana, E., Band, S. S. Comprehensive Review of Deep Reinforcement Learning Methods and Applications in Economics. Mathematics. 2020; 8:10, 1640
13. Zheng, S., Trott, A., Srinivada, S., Parkes, D. C., Socher, R. The AI Economist: Optimal Economic Policy Design via Two-level Deep Reinforcement Learning. arXiv. 2021.