“Deep knowledge” is David March’s proposed way to move beyond finding patterns in data: build a model of the underlying system, then examine how that system might behave when conditions change. His suggested approach pairs machine learning with an agent-based model. It is a conceptual proposal—not an established successor to deep learning or a method whose effectiveness the article demonstrates.
What does “deep knowledge” mean?
In his November 7, 2018 article, David March describes learning as acquiring or changing behavior or preferences, and knowledge as modifying or enhancing understanding. His distinction is meant to highlight a limitation he sees in machine learning: a model can reveal patterns in observed data without explaining the system that generated them.
That difference matters when conditions change. A model fitted to a narrow range of circumstances may capture what tends to happen there but miss hidden constraints, nonlinear effects, or feedback loops. If those underlying factors shift, its familiar patterns may no longer hold. March’s phrase “deep knowledge” names the desired understanding of how the system works and how it might respond to such changes; it is not a standardized technical stage in AI.
How does March propose moving from patterns to system understanding?
March suggests using machine-learning patterns as a target for an agent-based model. Such a model represents a system through agents—individual entities whose behaviors and interactions can combine to produce larger-scale, emergent patterns.
What’s actually slowing this PC down?
Pick the symptom - the matching free tool is one click away.
#1 Best Overall
- Language Published: English
- Binding: hardcover
- It ensures you get the best usage for a longer period
- Find patterns with machine learning. Identify recurring behavior in the observed data.
- Represent the system as agents. Define plausible agent behaviors and the equations or rules governing their interactions.
- Adjust the model. Change those behaviors and rules until the agents’ combined behavior resembles the patterns found through machine learning.
- Compare plausible configurations. Consider whether different underlying agent rules could produce similar observed patterns.
- Explore changed conditions. Use sensitivity analysis to examine how the modeled system might respond when relevant forces or constraints shift.
March summarizes the iterative modeling idea this way: “The strategy is to iteratively manipulate the parameters and equations that govern agent behavior until we are able to generate the emergent behavior that creates the same ML patterns.” That describes his proposed workflow; matching observed patterns does not, by itself, prove that the model has recovered the real system.
What does the customer-satisfaction example show?
March uses customer satisfaction as a thought experiment, not as a reported experiment or statistic. Imagine that machine learning groups customers who currently occupy similar positions in a satisfaction domain. Their apparent similarity might reflect a market force holding different responses in place. If that constraint were removed, customers’ behavior could diverge.
Rank #2
He invokes changes in interest rates and hyperbolic discounting to illustrate how a factor that appears unimportant or stable under current conditions might matter more after circumstances change. In an agent-based model, a researcher could represent customers with different behavioral rules, then explore how the modeled outcomes shift when the market force changes. The exercise can make assumptions visible and generate scenarios; it does not establish what real customers will do.
What are the limits of the proposal?
The approach depends on choosing agents, behaviors, equations, and constraints that adequately represent the real system. Similar patterns in observed data may be compatible with different underlying mechanisms, so reproducing those patterns alone cannot determine which explanation is right. Sensitivity analysis shows how a model responds to changed inputs or assumptions; it does not independently verify that those assumptions describe reality.
Recommended Free Tools
Rank #3
- Pattern prediction: Does the model forecast outcomes accurately under conditions like those represented in its data?
- Mechanism explanation: Are important variables and feedback loops represented in a plausible way?
- Changed conditions: Has the model been checked against observed behavior when conditions shift, rather than only explored in scenarios?
- Validation: What evidence supports the model’s assumptions and conclusions?
March’s article proposes a way to investigate these questions but does not report a validation study showing that the method reliably reconstructs real systems or outperforms alternatives. Its value should therefore be judged as a modeling idea to test in a particular domain, not as a proven general recipe.
Is agent-based modeling the next step for AI?
No single approach has been established as the next step after deep learning. A review published September 1, 2026, in Frontiers in Science discusses foundation models, generative AI, hybrid and neuro-symbolic architectures, and agentic AI as developing directions in medicine. It also emphasizes that real-world deployment requires validation, integration, safety work, and governance. That review is focused on medicine, so it offers a bounded example of current directions rather than a universal roadmap for AI.
Rank #4
March’s proposal addresses a narrower question: how might a researcher move from patterns learned from data toward an explicit model of the system that produced them? Agent-based modeling is one possible way to explore that question. It is not a settled replacement for deep learning, and the broader field continues along multiple paths.
Independent reader supportYour contribution helps us test, update, and keep practical guides available for everyone.When might this approach be useful?
It may be worth considering when the goal is to understand how interacting entities, hidden constraints, or feedback loops could generate an observed pattern—and when there is enough domain knowledge and evidence to build and test a meaningful model. If the immediate need is prediction within familiar conditions, a pattern-focused model may be more direct. If the question concerns mechanisms or changed conditions, an agent-based model can make assumptions explicit, but its conclusions still need validation against real-world evidence.
Quick Recap
Best Value
Product prices and availability are accurate as of the date/time indicated and are subject to change. Any price and availability information displayed on Amazon at the time of purchase will apply.
