Chapter 7: Related Theories and Neosis in Context

Over the past fifty years, multiple scientific communities have attempted to construct systems capable of evolving their computational structure, adapting to their environment, and developing new capabilities over time. Early work in the 1970s and 1980s explored evolutionary algorithms and genetic programming as mechanisms for open-ended problem solving [@holland1975adaption]. These systems demonstrated that mutation and selection could navigate immense design spaces, but the resulting artifacts were static: once deployed, their structures no longer changed within their lifetime.

During the 1990s and 2000s, Artificial Life (ALife) platforms such as Tierra and Avida [@ray1991tierra; @adami1998introduction] produced digital organisms that replicated, competed, and diversified under evolutionary pressures. These environments achieved open-ended population dynamics but failed to produce agents with rich internal models or capacity for structural reorganization within a lifetime. Evolution optimized replication, not adaptive cognition.

Parallel developments in neural networks and reinforcement learning [@lecun2015deep; @sutton2018reinforcement] yielded powerful functional systems operating over fixed network architectures. Neuroevolution methods such as NEAT and HyperNEAT [@stanley2002evolving; @stanley2009hyperneat] introduced structural evolution, but only across generations, not as part of an individual agent's adaptive cycle.

Recent work—neural cellular automata [@mordvintsev2020growing], hypernetworks [@ha2016hypernetworks], meta-learning, and developmental cognitive architectures—introduces forms of plasticity and structure manipulation. Yet none unifies in-lifetime structural change, survival-based objectives, and energy-constrained mutation into a single organismal model. Neosis aims to fill precisely this gap.

To situate Neosis clearly, this chapter presents:

1. A conceptual diagram mapping theoretical traditions into a three-axis space.

2. A categorical table comparing frameworks by objectives, and development.

3. A review of theoretical frameworks, organized into conceptual categories.

4. A review of systems and projects, organized into mechanistic categories.

Conceptual Placement of Neosis

Figure 7.1 — Conceptual Cube Diagram (Placeholder)

(A diagram to be inserted.)

Table 7.1 — Three-Axis Mapping of Frameworks

Theory Family	Objective Source	Developmental Regime
Artificial Neural Networks	External loss	Static
Spiking Neural Networks	External loss	Static
Reinforcement Learning	External reward	Static
Neuroevolution	External task fitness	Generational
Artificial Life (Tierra, Avida)	Replication fitness	Generational
Predictive Processing / FEP	Free-energy minimization	Static
Developmental Systems Theory	Biological fitness	Scaffolded
Dynamical Systems Theory	Emergent	Continuous
Embodied/Enactive Cognition	Viability	Developmental
Reservoir Computing	External training	Static
Neural Architecture Search	External loss	Static
Neural Cellular Automata	External constraints	Pattern-growth
Hypernetworks / Meta-learning	External loss	Static
Neosis	Internal survival energy	Open-ended

7.1 Theoretical Frameworks

Theoretical traditions can be grouped into categories based on the goal or approach shaping each field. These categories clarify how each tradition attempts to answer a subset of the problem Neosis addresses.

7.1.1 Category A — Learning-Based Theories

Goal: Improve performance through parameter adaptation while assuming a fixed architecture.

This category includes artificial neural networks, reinforcement learning, and predictive processing.

Artificial neural networks [@rumelhart1986learning; @lecun2015deep] use gradient-based updates to refine numerical parameters on static computation graphs. Reinforcement learning [@sutton2018reinforcement] optimizes expected reward by adjusting policy or value function parameters based on experience. Predictive processing and the Free Energy Principle [@friston2010free] treat cognition as hierarchical prediction-error minimization.

Despite their strengths, these learning-based theories rely on fixed structures and externally defined tasks. Their adaptation occurs exclusively through differentiable updates rather than through structural mutation. Because they do not couple prediction, reward, and structural change within a single organismal cycle, they lack the unified energy economy and continuous self-modification that characterize Neosis.

7.1.2 Category B — Evolution-Based Theories

Goal: Discover functional structures through mutation and selection across generations.

Evolutionary computation [@holland1975adaption], genetic programming [@koza1992genetic], and neuroevolution methods such as NEAT and HyperNEAT [@stanley2002evolving; @stanley2009hyperneat] search vast design spaces through generational mutation and selection. Artificial Life environments such as Tierra and Avida [@ray1991tierra; @adami1998introduction] demonstrate open-ended population dynamics and developmental divergence.

Although these systems generate novelty and support domain-general search, they do not allow in-lifetime structural change. Adaptation is tied to replication fitness rather than predictive survival. Internal cognitive complexity remains limited because there is no mechanism linking computation, reward, and structural modification inside the organism's operational cycle. Neosis unifies learning and evolution in a way these frameworks do not.

7.1.3 Category C — Developmental and Cognitive Theories

Goal: Explain how cognition emerges from developmental processes, modularity, and environmental interaction.

Computational cognitive science [@anderson2007integrated; @newell1994unified] models cognition through specialized modules and multi-timescale adaptation. Developmental systems theory [@oyama2000ontogeny] emphasizes gene–environment coupling and emergent developmental trajectories. Embodied and enactive cognition [@varela1991embodied; @clark1997being] argue that cognition arises from tight coupling between organism and environment.

These theories articulate powerful principles describing natural cognition, yet they are interpretive rather than constructive: they explain biological systems rather than define a minimal computational substrate for synthetic organisms. Because they lack explicit mutation mechanisms or survival-based internal objectives, they do not exhibit the energy-constrained structural evolution that Neosis supports.

7.1.4 Category D — Self-Organizing and Dynamical Theories

Goal: Reveal how complex behavior emerges from simple local interactions and intrinsic system dynamics.

Dynamical systems theory [@kelso1995dynamic] models cognition and behavior as trajectories through attractor landscapes shaped by system interactions. Reservoir computing [@jaeger2001echo] uses fixed recurrent dynamics to produce rich transformations, relying on trained readouts. Neural cellular automata [@mordvintsev2020growing] demonstrate pattern formation and self-repair based on local update rules.

Although these systems produce rich emergent dynamics, they provide no survival objective and no mechanism for energy-regulated mutation. Most operate on fixed architectures or predetermined update laws. Without a unified operational cycle linking prediction, reward, structural change, and survival, their adaptability remains fundamentally limited compared to Neosis.

7.2 Systems and Projects

Systems and projects differ from theoretical frameworks in that they implement working platforms. These can likewise be grouped into categories based on their underlying mechanistic approach.

7.2.1 Category E — Evolutionary ALife Platforms

Approach: Create digital ecosystems where organisms mutate, replicate, and compete.

Platforms such as Tierra [@ray1991tierra], Avida [@adami1998introduction], Polyworld [@yaeger1994computational], and Lenia [@chan2019lenia] explore ecological competition, mutation, and population-level adaptation. These systems showcase open-ended dynamics, spontaneous diversification, and ecosystem-level complexity.

Yet despite their ecological richness, these systems rarely develop sophisticated internal predictive models or meaningful within-lifetime structural adaptation. Their objectives remain tied to replication rather than survival-based computation, and mutation affects only generational change, not continuous structural reorganization. Neosis's integration of prediction, energy, and mutation within a single organism sets it apart.

7.2.2 Category F — Adaptive Neural Systems

Approach: Enhance neural models with auxiliary mechanisms for dynamic parameterization or meta-level adaptation.

Hypernetworks [@ha2016hypernetworks], meta-learning architectures, differentiable architecture search, and morphological computation [@pfeifer2007self] provide additional flexibility beyond fixed neural parameters. They introduce mechanisms that generate weights dynamically or reconfigure computational pathways.

However, these systems maintain fixed meta-structures and rely on external, differentiable losses to govern adaptation. They do not incorporate intrinsic survival-based objectives or unified energy economies that regulate structural mutation. Consequently, they lack the open-ended structural growth that Neosis enables within a single lifetime.

7.2.3 Category G — Self-Modifying Neural or Graph Systems

Approach: Allow networks to alter their topology or computation rules during operation.

Neural cellular automata [@mordvintsev2020growing], continual topology-adapting networks [@stanley2003continual], and modular neural architectures introduce forms of structural plasticity that operate during computation. These approaches explore the frontier between fixed computation graphs and flexible, evolving structures.

Despite exploring structural change, these systems lack a survival-based internal energy model that ties structural mutation to adaptive success. Structural modifications are typically driven by heuristics, external training signals, or rule-based mechanisms rather than a unified organismal loop linking prediction, reward, and survival. As such, they do not achieve the open-ended cognitive development central to Neosis.

7.3 Lessons for Micro and Macro Neosis from Related Theories and Systems

Micro-Neosis: Goal, Approach, and Design Lessons

Micro-Neosis aims to define the fundamental computational laws governing a single Neo: how binary nodes update, how energy is spent, how prediction is computed, and how structure modifies itself autonomously within a lifetime. Its approach is to treat each Neo as a minimal, nondifferentiable, energy-regulated graph whose dynamics arise entirely from local Lex rules, Evo mutation primitives, and the internal economy of Nex. From related work, Micro-Neosis can meaningfully adopt concepts such as local state-update stability from dynamical systems, noise-aware computation from neural and stochastic models, mutation operators from evolutionary computation, and multi-timescale adaptation principles from developmental theory. At the same time, it must avoid the limiting assumptions that caused prior systems to stagnate: fixed architectures as in ANN and RL; evolution tied purely to replication rather than cognition as in ALife; structural plasticity that is unconstrained or purposeless as in many self-organizing systems; and reliance on differentiability, which restricts open-ended modification. The micro architecture succeeds precisely by integrating structural mutation, computation, and energy into one closed organismal loop—something earlier systems did not achieve.

Macro-Neosis: Goal, Approach, and Design Lessons

Macro-Neosis aims to construct a coarse-grained mathematical representation of a Neo that captures its global input–output behavior, stability properties, specialization tendencies, and emergent cognitive structure without simulating individual nodes. Its approach uses linearization, stochastic perturbation theory, controllability/observability concepts, and filtering-based abstractions to compress high-dimensional micro dynamics into tractable macro equations. From related work, Macro-Neosis can draw heavily on control theory for stability analysis, on nonlinear filtering for representing aggregated noise from micro-level stochasticity, on reservoir-style approximations for dimensionality reduction, and on developmental and cognitive theories for understanding functional specialization. Yet it must avoid the pitfalls common in these traditions: reliance on smooth differentiable models that fail under discrete stochastic updates; macro abstractions disconnected from evolving structure; the absence of grounding in an explicit micro substrate; and dependence on brute-force simulation, which plagued many ALife projects. Macro-Neosis retains the strengths of these fields but overcomes their weaknesses by ensuring that its coarse models remain dynamically consistent with micro-level Lex, Evo, and Nex dynamics.

7.4 Summary

Neosis draws from decades of research in neural computation, Artificial Life, evolutionary algorithms, cognitive science, and self-organizing systems. Prior efforts focused on learning without structural evolution, evolution without cognition, or development without energy constraints. Neosis unifies these principles into a single minimal computational framework linking prediction, energy, structure, and survival. This positions Neosis not as an incremental extension of prior systems, but as a foundation for open-ended, self-modifying digital organisms.