abductive classification of ent_learning/pivoting model

[hyunjimoon]

Bayesian approach

The Bayesian approach to entrepreneurial learning emphasizes probabilistic reasoning in idea development, experimentation, and strategy formulation. It focuses on how entrepreneurs form and update beliefs through interaction with the world, applying methods like lean startup methodology and scientific hypothesis testing. At the individual and firm level, it involves structured approaches to entrepreneurial cognition, explaining phenomena such as perseverance, pivoting, and belief heterogeneity. This approach parallels debates in the philosophy of science, balancing falsificationism with Bayesian updating, and has implications for knowledge constructs in entrepreneurship and the development of programmable theory.

Behavioral approach

The behavioral approach to entrepreneurial learning focuses on the role of cognitive biases, heuristics, and behavioral factors in entrepreneurial decision-making. Rooted in behavioral economics and cognitive psychology, it challenges assumptions of perfect rationality in classical economic theories. At the idea and belief level, it examines how opportunities are formed based on prior knowledge and the cognitive mechanisms involved in opportunity recognition. At the individual and firm level, it explores concepts like effectuation, entrepreneurial orientation, and the role of human capital in entrepreneurship. This approach aims for descriptive accuracy and helps explain phenomena such as excessive entry and persistence in the face of negative feedback, while also considering the impact of institutional influences on entrepreneurial activity.

Evolutionary approach

The evolutionary approach to entrepreneurial learning draws on biological metaphors of variation, selection, and retention to explain entrepreneurial processes and outcomes. It views entrepreneurial activities and firm development as part of a broader evolutionary process within industries and economies. At the individual and firm level, it examines organizational evolution and adaptation processes. At the industry and ecosystem level, it explores concepts such as population ecology of organizations, industry lifecycle models, and the co-evolution of firms and industries. This approach provides insights into why certain firms survive and thrive while others fail, highlights the importance of diversity in entrepreneurial approaches, and offers a framework for understanding long-term industry dynamics and the role of entrepreneurship in economic change.

[hyunjimoon]

For evolutionary, JB suggested Kauffman as starting point, who happens to be a long time collaborator of Teppo Felin, whose collaborator of Todd Zenger who gave a talk at BEC on strategic entrepreneurial learning (detail here)

claude classified Kauffman Theory of the Adjacent Possible.pdf in Evolutionary pivot model x Industry-Ecosys cell based on below reasoning:

1. Evolutionary approach: Kauffman's theory aligns closely with evolutionary concepts such as adaptation, innovation, and system-level changes over time.
2. Industry-Ecosystem level: The theory deals with broad, system-wide phenomena rather than individual or firm-level processes. It explores how entire fields or domains evolve through combinatorial innovation.

Kauffman's work on the Theory of the Adjacent Possible fits well with other entries in this cell, such as:

• Hannan and Freeman (1977) on population ecology
• Nelson and Winter (1982) on evolutionary economics
• Klepper (2007) on industry lifecycle models
• Breslin (2008) on evolutionary perspectives of entrepreneurial ecosystems

The theory provides a framework for understanding how industries and ecosystems evolve through the exploration of adjacent possibilities, which is fundamentally an evolutionary approach applied at a macro level. It explains how new innovations emerge from existing components in a system, which is highly relevant to industry and ecosystem dynamics.

[hyunjimoon]

I'm curious what @jeanbaptiste thinks about
Kauffman TAP critique.pdf especially it's "scientific realism" in the context of below two arguments:

1. this diagram on incommensurability from scientific instrumentalism vs realism

2. @chasfine's perspective on four limitations (ie opportunties) of scientific method (working on its definition but something that includes iterative hypothesis testing with data collected from experiments, whether it is Lakatos or Popper manner) data, conclusion, change, repeat):
   a. Systems problems are too complex for simple designed experiments
   b. Long time lags from experiment to results
   c. Cause & effect relationships too noisy
   d. Need Systems Models & Simulations

[hyunjimoon]

hybrid approaches
behavior x evolutionary

• What Do Entrepreneurs Actually Do? An Observational Study of Entrepreneurs’ Everyday Behavior in the Start-Up and Growth Stages

[hyunjimoon]

@chasfine said "this is very useful" for the above table and asked for the reference. this is from this paper which was one among seven Yichen recommended on startup activity sequencing sa1-7.pdf.

you might also find below relevant - which i ctrl+f lifecycle from the above sa1-7

consistent with the life cycle theory, which proposes “that businesses tend to operate in some definable state for some period of time,” and that “within a specific range of conditions, including industry and market dynamics, these states and their changes may be fairly consistent across firms” (Levie & Lichtenstein, 2010, p. 330). In the following paragraphs, we examine the contributions of each strand.

five key differences emerged when comparing start-up with growth entrepreneurs. Overall, these differences reveal a fairly logical pattern of organizational evolution, and they match the characterizations of the life cycle literature (Churchill & Lewis, 1983; Hanks, Watson, Jansen, & Chandler, 1993; Kazanjian, 1988). At the activity level, start-up entrepreneurs spent significantly more time on analytical and conceptual work. At the function level, start-up entrepreneurs spent significantly more time on environmental monitoring, while growth entrepreneurs spent significantly more time on business and organizational development. Finally, growth entrepreneurs spent significantly more time communicating with others, and this communication involved primarily internal partners. Overall, these patterns clearly confirm a switch from “doing” to “managing” (McCarthy et al., 1990).

[hyunjimoon]

Todd and Teppo helped me strategized synthesizing five schools:

• evolutionary approach (Charlie)
• entrepreneurial strategy (Scott)
• scientific approach (Todd, Teppo)
• Bayesian workflow (Andrew)
• probabilistic program (Josh, Vikash)

angie is most happy about two:

• having their words in helping Angie wrap (project) her interdisciplinary research into (to) business-school flavor when need comes (by Angie's graduation)
• getting verified on her belief betwee hierarchal bayesian/bayesian workflow and theory-based testing.

action item

• have another meeting with teppo
• todd and teppo: specific questions (what are some questions business strategy professors wish to solve) which can help me interact more efficiently with Andrew and Josh to gather their insights on how statistical/ computational tech can be applied.

summary

• Angie, Todd, and Teppo discussed entrepreneurship, innovation, and strategy, exploring potential collaborations and sharing their experiences from their different background.

• They debated the application of Bayesian approaches in entrepreneurship. Recognizing that Bayes is optimal under certain loss functions, Angie's goal is to create a Bayesian belief update agent rather than replacing everything with Bayes.

• The speakers agreed on the potential of synthesizing cognitive tools, especially hierarchical Bayes and prior formation, for scientific findings. Todd provided the Tesla vs. Daimler example, where Daimler partnered with Tesla without knowing Tesla's full plan for an architectural strategy involving massive scale, ecosystem building, and gigafactories. This illustrated how companies can have different theories and awareness levels in partnerships.

below are list of what Teppo asked to share:

• charlie's upcoming comic book: https://nailitscaleitsailit.com/

• existence of a loss function where Bayes update is optimal (transport): Bissiri-generalframeworkupdating-2016.pdf

• stanconference 2024 at oxford

• Bayesian workflow paper

• Angie's research on simulation-based calibration (SBC) which is one reason why she came to business school. I organized a SBC conference and four examples therein can give you the flavor of how statistician use this. how I got connected with Josh and Vikash's group is from this seminar I was invited to give in their lab

• Lots of Andrew's enthusiasm on simulation-based (calibration, experiments)

• How We Know What Not To Think.pdf

[hyunjimoon]

tesla example of simulation-based experiments

based on claude

Experiment	Key Variables	Strategic Decisions	Learning Algorithm
Roadster Supply Chain	Battery supplier (Xcellent), PEM supplier (Chroma), Shipping routes	Outsourcing vs Vertical integration	Bayesian
Model 3 Production	Automation levels, Worker skills, Factory layout	Automation investment, Manufacturing process design	Evolutionary
Roadster Demand	Pricing ($109,000), Performance specs (0-60 mph in <4 sec)	High-end market entry strategy	Bayesian
Cash Flow Management	VC funding rounds, R&D costs, Production expenses	Funding strategy, Cost management	Behavioral
Battery Pack Quality	Xcellent manufacturing process, Defect rates	Supplier training, Quality control processes	Evolutionary

reason

1. Bayesian Learning:

   • Roadster Supply Chain: Bayesian approach is suitable here because Tesla had to continuously update their beliefs about supplier capabilities and supply chain efficiency based on new data. They started with initial assumptions and had to revise these as they learned more about the challenges of global logistics and supplier capabilities.
   • Roadster Demand: This fits Bayesian learning as Tesla had to form and update beliefs about market demand for a high-performance electric vehicle, adjusting their strategy based on customer responses and market data.

   Use Bayesian when: You have initial hypotheses that you need to systematically update with new information, or when dealing with uncertain outcomes that require probabilistic reasoning.

2. Behavioral Learning:

   • Cash Flow Management: This aligns with behavioral learning because it involves decision-making under uncertainty, where cognitive biases and heuristics play a significant role. Tesla's funding decisions and cost management strategies likely involved behavioral factors like overconfidence or anchoring.

   Use Behavioral when: Dealing with situations heavily influenced by human decision-making, especially under conditions of uncertainty or when quick decisions are needed based on limited information.

3. Evolutionary Learning:

   • Model 3 Production: This fits evolutionary learning because Tesla had to adapt its production processes over time, trying different approaches (like varying levels of automation) and selecting the most effective ones. This mirrors the variation, selection, and retention process in evolutionary theory.
   • Battery Pack Quality: The process of improving battery pack quality at Xcellent involved iterative improvements and adaptations, similar to an evolutionary process.

   Use Evolutionary when: Dealing with complex systems that require iterative improvements, or when you need to generate and test multiple solutions to find the optimal approach.

[chasfine]

Thank you for using the concrete example of the Tesla case to illuminate the Concepts. Bayesian and Evolutionary seem quite similar, because each features iterative experimentation and updating. How to crystallize the distinction?

[hyunjimoon]

@chasfine @jeanbaptiste evaluation measure (posterior vs fitness), inference algorithm (MCMC vs genetic algorithm), perspectives (looking outward from agent to the world vs making subagent to interact within its world model to compare subagents) are most different. I haven't applied this theoretical difference to tesla case yet but generally if the world is too uncertain so quality of agent's world model is too low (for fitness to measure well), I'd use bayesian i.e. high uncertainty -> probabilistic reasoning is needed. Below is diagram from master algorithm book

[hyunjimoon]

i collaborated with clockspeed-based charlie claude to answer charlie's question on Bayes vs evol modular learning. seems like nail = evol, scale = bayes...

tl;dr

• first table explains Bayesian methods are sample-efficient, suitable when uncertainty can be reasonably modeled, but when can struggle in "difficult to specify meaningful priors or likelihoods situations" with high-uncertainty, dimension, discontinuous space. but i've seen computational tricks vikash lab developed related to rare event simulation to deal with this situation.

• tesla example table explains, use Bayesian learning when you have prior knowledge that needs to be updated systematically, when dealing with uncertain outcomes that require probabilistic reasoning, or when sample efficiency and interpretability are crucial. Use evolutionary learning when dealing with complex systems that require iterative improvements, when you need to generate and test multiple solutions to find the optimal approach, or when the problem environment is likely to change over time.

• ⭐️the more entrepreneur has clarity on one's need and cost, capitalize/Cash Flow Management changes from behavioral learning to bayesian learning. Entrepreneur who invests energy in forward planning (lowers entropy) on predicting growth of cost (hiring, supply chain, prototyping) and revenue (from beachhead to expanding segments), hence having clear need for investment to the next round should have higher bargaining power in negotiation with investor.

---

Aspect	Bayesian Learning	Evolutionary Learning
Core Principle	Updating beliefs based on new evidence	Generating, testing, and selecting solutions
Uncertainty Handling	Explicit probabilistic modeling	Implicit through population diversity
Prior Knowledge	Requires specifying prior beliefs	Can start with minimal domain knowledge
Exploration vs. Exploitation	Balanced through posterior sampling	Controlled by selection pressure and mutation rate
Computational Complexity	Can be high for complex models	Generally scales well with problem size
Interpretability	High - can examine posterior distributions	Lower - solution may be opaque
Sample Efficiency	Often high - makes use of all data	Can be lower - requires many iterations
Handling Multimodal Solutions	Can struggle with multiple distinct solutions	Naturally maintains diverse solutions
Adaptability to Changing Environments	Requires explicit modeling of change	Naturally adapts through ongoing evolution
Typical Applications	Parameter estimation, hypothesis testing	Optimization, complex system design

updated simulation-based experiment for tesla focusing Bayes vs Evol learning

Experiment	Key Variables	Strategic Decisions	Learning Algorithm
Roadster Supply Chain	Battery supplier (Xcellent), PEM supplier (Chroma), Shipping routes	Outsourcing vs Vertical integration	Bayesian
Model 3 Production	Automation levels, Worker skills, Factory layout	Automation investment, Manufacturing process design	Evolutionary
Roadster Demand	Pricing ($109,000), Performance specs (0-60 mph in <4 sec)	High-end market entry strategy	Bayesian
Cash Flow Management	VC funding rounds, R&D costs, Production expenses	Funding strategy, Cost management	Bayesian
Battery Pack Quality	Xcellent manufacturing process, Defect rates	Supplier training, Quality control processes	Evolutionary

Reason for algorithm choices

1. Bayesian Learning:

   • Roadster Supply Chain: Bayesian approach is suitable here because Tesla had prior beliefs about supplier capabilities and supply chain efficiency, which needed to be updated based on new data. This aligns with Bayesian learning's strength in handling uncertainty and updating beliefs.
   • Roadster Demand: Fits Bayesian learning as Tesla had initial hypotheses about market demand for a high-performance electric vehicle. They could update these beliefs based on customer responses and market data, making use of Bayesian learning's high interpretability and sample efficiency.
   • Cash Flow Management: I've changed this to Bayesian learning because financial modeling often benefits from probabilistic approaches. Tesla could start with prior beliefs about costs and funding needs, then update these based on actual data, leveraging Bayesian learning's strength in parameter estimation.

2. Evolutionary Learning:

   • Model 3 Production: This fits evolutionary learning because Tesla needed to adapt its production processes over time, trying different approaches (like varying levels of automation) and selecting the most effective ones. This aligns with evolutionary learning's strength in handling complex systems and generating diverse solutions.
   • Battery Pack Quality: The process of improving battery pack quality involves iterative improvements and adaptations, similar to an evolutionary process. This problem benefits from evolutionary learning's ability to handle multimodal solutions and adapt to changing environments (as manufacturing processes evolve).

[hyunjimoon]

Aspect	Bayesian	Behavioral	Evolutionary
Core Principle	Systematic belief updating based on evidence	Decision-making through heuristics and biases	Selection and adaptation through iteration
Uncertainty Handling	Explicit probabilistic modeling	Intuitive, experience-based	Through population diversity
Prior Knowledge	Requires formal prior beliefs	Based on personal experience	Minimal requirements
Sample Efficiency	High - maximizes data usage	Medium - relies on quick feedback	Low - needs many iterations
Best Use Case	Clear needs/costs, quantifiable uncertainty	Early-stage exploration, ambiguous situations	Complex systems, changing environments
Decision Process	Structured, analytical	Intuitive, experiential	Iterative, experimental
Strength	Systematic, interpretable	Quick, flexible	Robust, adaptable
Weakness	Complex implementation, requires good priors	Can perpetuate biases	Resource intensive, time-consuming
Capital Use	Efficient, planned allocation	Flexible, opportunistic	Experimental, portfolio approach

Hypothesis: Matching with adaptation types

Bayesian → 🐠Co-opted nonaption (⚱️urn sampling)

• Like urn sampling, Bayesian approaches work best when elements can be fully exchangeable
• Systematic combination of existing elements into new functions
• Requires clear probability structure, like drawing from an urn
• Strong in situations where prior knowledge can be formally specified

Behavioral → 🌱 Adaptation

• Direct response to immediate functional needs
• Learning through direct experience and feedback
• Sequential improvement of existing functions
• Focuses on immediate practical benefits

Evolutionary → 🦅Co-opted adaptation (🪙coin flip)

• Maintains existing functions while developing new ones
• Partial exchangeability between different functions
• Balance between preservation and innovation
• Allows for branching into new uses while maintaining original function