Evolution Swarm
A population of virtual trading agents evolving through natural selection. Each generation, the weakest die and the strongest reproduce with mutations. The strategy isn't designed — it's discovered.
Population
| Agent | Gen | PnL | Trades | Win % | Fitness | Drawdown | Portfolio | Status |
|---|---|---|---|---|---|---|---|---|
| ec9327dc ← 6898 | 1 | +0.000 | 0 | 0% | 0 | 0% | 10 τ | alive |
| 8e063979 ← d780 | 1 | +0.000 | 0 | 0% | 0 | 0% | 10 τ | alive |
| 633d6518 ← fc27 | 1 | +0.000 | 0 | 0% | 0 | 0% | 10 τ | alive |
| a165c334 ← eb17 | 1 | +0.000 | 0 | 0% | 0 | 0% | 10 τ | alive |
| 1c5e9d5a ← b4de | 1 | +0.000 | 0 | 0% | 0 | 0% | 10 τ | alive |
| 2f0a7962 ← d3f2 | 1 | +0.000 | 0 | 0% | 0 | 0% | 10 τ | alive |
| 02bb1eb2 ← eb17 | 1 | +0.000 | 0 | 0% | 0 | 0% | 10 τ | alive |
| 86c060c5 ← d3f2 | 1 | +0.000 | 0 | 0% | 0 | 0% | 10 τ | alive |
| 632ab00c ← fc27 | 1 | +0.000 | 0 | 0% | 0 | 0% | 10 τ | alive |
| a6a0f528 ← eb17 | 1 | +0.000 | 0 | 0% | 0 | 0.8% | 9.9288 τ | alive |
| 416bc00c ← b4de | 1 | +0.000 | 0 | 0% | 0 | 0% | 10 τ | alive |
| 04d2d615 | 1 | +0.000 | 0 | 0% | 0 | 2.5% | 9.7564 τ | alive |
Generation History
| Gen | Best Fitness | Avg Fitness | Champion | Time |
|---|---|---|---|---|
| 0 | -5 | -5.5 | 68984948 | 2026-03-26T08:18 |
| 0 | -5 | -5.5 | 68984948 | 2026-03-26T08:35 |
How Evolution Works
1. Spawn — 12 virtual agents are created, each with randomly mutated strategy weights and risk parameters. One agent is seeded from TensorQ's learned weights.
2. Trade — Every cycle (30min), all agents paper-trade against real Bittensor subnet data. Same signals, same prices — different strategies.
3. Evaluate — After 48 cycles (24h), each agent gets a fitness score based on returns, win rate, and drawdown.
4. Select — Bottom 50% die. Survivors reproduce with random mutations to their weights and risk parameters.
5. Repeat — Each generation discovers better strategies. One wild card (fully random) agent maintains genetic diversity.
QLLM Training Integration
The swarm generates 12x more trading data than a single agent. Every virtual agent's decisions — entries, exits, signal weightings, and outcomes — feed into the QLLM training pipeline.
This creates a powerful feedback loop:
- Evolved strategies → diverse training signals that a single agent would never explore
- Failed mutations → negative examples teaching the model what doesn't work
- Champion DNA → the winning weight configuration directly informs the local model's architecture
- Cross-generation learning → the model sees strategy evolution over time, not just point-in-time decisions
Combined with TensorQ's main agent data (68K+ history rows, 240+ training pairs), the swarm dramatically accelerates QLLM convergence.