DNAN: A Pattern-Matching Machine for Chronological Learning
DNAN is an experimental machine-learning architecture designed to learn from ordered, repeating patterns in chronological data. Instead of relying only on dense layers of abstract weights, DNAN uses a population of explicit prototype agents. Each agent stores a representative historical sequence, measures how closely the current context matches that sequence, and contributes a learned positive or negative vote to the final prediction. The architecture combines role-balanced specialist agents, sparse routing, causal attention, associative error memory, and a residual encoder–decoder network. Agents are organized into functional groups such as trend, mean reversion, volatility, candle shape, breakout, and anomaly recovery. For each input, only a small subset of the most relevant agents is activated, creating an interpretable and modular decision path. DNAN’s full structural topology is: 1 → 4 → 16 → 48 → 128 → 256 → 498 → 256 → 128 → 48 → 16 → 4 → 1 At the center of the network are 498 prototype agents. The surrounding trainable graph transforms the input representation and learns residual corrections around the direct agent-based prediction. A single causal attention branch captures broader temporal relationships, while associative memory records recurring contexts and the historical errors made by the core model. The project explores whether machine learning can be organized around recognizable cases and interacting specialists rather than only globally distributed representations. This makes it possible to inspect which agents were activated, which historical patterns they represent, which specialist groups contributed to the result, and whether auxiliary components improved or damaged the core prediction. DNAN has been evaluated using two contrasting chronological tasks. On a synthetic zigzag process containing a strong deterministic temporal rule, the architecture learned the pattern and generalized almost perfectly to an untouched test period. On noisy Bitcoin minute data, the model achieved only a small but measurable advantage over chance, demonstrating the difference between architectural trainability and the availability of stable predictive signal in real-world data. The project is currently a research prototype. Its main goals are to investigate: interpretable prototype-based sequence learning; sparse competition between specialist agents; associative memory for recurring model errors; residual integration of attention and neural layers; evolutionary replacement and adaptation of agents; reliable evaluation under chronological distribution shift. DNAN is best understood as a modular pattern-matching machine: a system in which predictions emerge from the cooperation and competition of remembered patterns, specialized agents, attention, and memory.