Oncomorphic neural agent populations for resource-limited sequential learning

Technical Summary: Oncomorphic Neural Agent Populations for Resource-Limited Sequential Learning

Problem Statement
Distributed artificial intelligence systems often operate under constraints of sequential task exposure, uneven computational resources, and decentralized coordination. Traditional continual learning approaches, such as parameter regularization or replay strategies, typically assume a centrally managed computational budget and evaluate candidate models outside a spatially distributed ecology. However, real-world edge devices, sensor networks, and robotic collectives face data streams that arrive locally, tasks that change over time, and compute budgets that vary by location. The paper addresses the challenge of determining which neural network architectures survive repeated exposure to heterogeneous tasks in a decentralized setting where variation, selection, mobility, and finite local support interact without centralized control.

Methodology
The authors propose an oncomorphic framework, a computational model inspired by cancer ecology, to simulate a decentralized population of neural agents. The system operates on two coupled scales:

1. Within-Agent Learning: Each agent carries a single dense feed-forward neural network and trains locally on the specific nonlinear classification task of its current environment using gradient descent. Training duration is finite and proportional to the environment's local carrying capacity ($K_e$), ensuring agents operate under bounded learning budgets rather than asymptotic convergence.
2. Between-Agent Ecological Turnover: The population evolves through four mechanisms:
   • Replication: Agents replicate at rates dependent on local performance and environment occupancy (logistic growth).
   • Mutation: Agents can mutate their architecture (depth, width, activation function, optimizer) to generate new designs.
   • Migration: Agents move between environments, potentially triggering the recruitment of carrying capacity from a finite shared reserve ($K_{res}$).
   • Selection: Poorly performing agents are removed based on a performance-dependent probability function.

The study utilizes controlled synthetic nonlinear classification environments (including argmax, Gaussian-blob, and orthant tasks). Two initial conditions were tested: initRand (a diverse population of random architectures) and initMax (a homogeneous population of maximal architectures). The authors performed parameter sweeps over selection strength ($\alpha$) and mutation rate ($\mu$), as well as recruitment probability and baseline replication rates, using a continuous-time Markov chain (Gillespie algorithm) for simulation.

Key Contributions

• Oncomorphic Framework: Introduces a novel multi-agent system where adaptation emerges from ecological turnover (migration, mutation, selection) rather than centralized optimization, explicitly modeling the interaction between local learning and finite ecological carrying capacities.
• State-Dependent Role of Mutation: Demonstrates that the utility of architectural mutation is contingent on the initial diversity of the population.
• Decentralized Multi-Task Competence: Shows that a population trained only on local, sequential tasks can, through ecological selection, become enriched with architectures that exhibit higher competence across the entire task set without additional joint training.
• Resource Redistribution vs. Performance: Clarifies that while recruitment and replication reshape demographic support, they act as redistribution mechanisms for finite resources rather than drivers of improved prediction quality.

Results

• Selection and Local Accuracy: Stronger performance-dependent selection consistently increased the terminal local accuracy of surviving agents. In the initRand condition (diverse start), high selection with low mutation yielded the best results (~0.878 accuracy), as selection consolidated existing high-performing architectures. In the initMax condition (clonal start), mutation was essential; without it, selection could not act on variation. High selection combined with nonzero mutation improved accuracy from ~0.48 to ~0.834.
• Multi-Task Competence: Selection on local performance correlated with improved end-of-run multi-task accuracy (evaluating agents on all environments without retraining). In initRand, multi-task accuracy peaked near 0.65 with high selection and low mutation. In initMax, mutation was again required to generate alternatives for selection to filter, reaching ~0.57.
• Recruitment Dynamics: Varying recruitment probability and baseline replication rates resulted in local accuracy remaining within a narrow band (0.785–0.853). This indicates that intensified demographic activity redistributes finite learning resources but does not systematically improve prediction quality.
• Time-Resolved Dynamics:
  • Concentration: Performance-dependent selection drove a significant concentration of the population toward a restricted set of architectures (dominant-architecture fraction increased to 0.23), whereas neutral drift produced only modest concentration (0.08).
  • Diversity: In clonal starts, mutation increased Shannon entropy (diversification), which selection then filtered. In diverse starts, mutation slightly reduced entropy by perturbing favorable designs.
  • Finite-Cycle Learning: Agents continued to improve throughout local training episodes and were evaluated before full convergence, suggesting that trainability under short windows is a critical factor for survival.
• Architecture Characteristics: The surviving architectures were often relatively compact (e.g., 2 hidden layers, 140 nodes) rather than the maximal depth/width initially present, indicating that efficiency under bounded budgets is favored over raw capacity.

Significance and Claims
The paper presents proof-of-concept mechanistic evidence that oncomorphic population dynamics offer a viable route to decentralized adaptation in engineering applications with bounded local resources. The authors claim that:

1. Useful competence in distributed systems can arise from population-level enrichment of architectures that perform well under sequential local exposure, rather than solely from preserving knowledge in a single persistent learner.
2. The effectiveness of exploratory pressure (mutation) is state-dependent: it is crucial when diversity is low (clonal starts) but potentially disruptive when a broad, favorable distribution of designs already exists.
3. There is a critical distinction between adaptation (selection and mutation determining which architectures persist) and allocation (recruitment and replication reshaping the demographic substrate).
4. Under finite training budgets, architectures that are "trainable" and efficient may outperform those with maximal theoretical capacity.

The authors explicitly note limitations, including the use of synthetic task families and architecture-level inheritance (without weight inheritance), framing the results as mechanistic evidence rather than a complete solution for real-world continual learning. They suggest future work could extend the framework to other agent types (CNNs, transformers) and incorporate explicit cost models for energy and latency.