Ant colony: a pheromone trail per tier, sampled against a delay heuristic

What it shows

Which tier earns the load is not fixed, it is negotiated by the traffic itself. AntColonyOrchestrator keeps one pheromone trail per tier (Mist, Edge, Cloud), seeded at INITIAL_PHEROMONE = 0.5. In find_computing_node a tier is sampled with a weight that multiplies the trail by a delay heuristic, weight = pheromone ** ALPHA / (1 + delay) ** BETA with ALPHA = 1.0 and BETA = 2.0, so a fast tier is preferred from the first task while the trail still adapts. In results_returned the trail of the tier that ran the task is updated as (1 - EVAPORATION) * P + EVAPORATION * quality with EVAPORATION = 0.1, where quality is max_latency / (max_latency + total_delay) for a success and 0 for a failure, and the result is clamped between MIN_PHEROMONE = 0.05 and MAX_PHEROMONE = 5.0. Those bounds are the point: a failed tier never drops to zero, so it stays a candidate, and a lucky tier never saturates, so one early win cannot lock the choice. Inside the chosen tier the node with the smallest estimated delay, counting the tasks already pending on it, executes the task.

flowchart TD
    task["Task"] --> best["_best_node per tier"]
    best --> weight["weight = pheromone^1.0 / (1+delay)^2.0"]
    weight --> sample["_sample over the three tiers"]
    sample --> run["Run on Mist, Edge or Cloud"]
    run --> result{"task SUCCESS?"}
    result -->|yes| q["quality = lat / (lat + delay)"]
    result -->|no| z["quality = 0.0"]
    q --> trail["P = 0.9*P + 0.1*quality"]
    z --> trail
    trail --> clamp["clamp to 0.05 .. 5.0"]
    clamp --> best
    classDef trail_style fill:#cfe8ff,stroke:#333;
    class trail,clamp trail_style;

The scenario

settings/ runs one scenario, once, on a fixed seed:

• 300 devices on a 2000 by 2000 meters map, with a 20 meters device-to-device (mist) range, on the ALL architecture and the ANT_COLONY algorithm, random_seed = 29, five simulated minutes, decisions made on the device (enable_orchestrators = false);
• a fleet of five device types, two of which are sensors with no computing power (0 cores), so only about 40% of the devices can host a mist task;
• four edge data centers of 10 cores at 40000 MIPS, one cloud of 200 cores, and a shared 20 Mbps WAN uplink with a 0.06 s latency;
• three applications: Health (Hard Real-time, 20%, 500 MI, 0.02 s bound), Augmented reality (Soft Real-time, 30%, 5000 MI, 0.5 s bound) and HEAVY_COMP_APP (Non Real-time, 50%, 30000 MI, 300 s bound).

Run it

From the repository root, after pip install -e .:

python examples/ant_colony/ant_colony.py

No-install variant:

PYTHONPATH=src python examples/ant_colony/ant_colony.py

What to look for

ANT_COLONY               devices= 300 tasks= 16870 ok= 12352 avg_delay=0.1666   cloud= 2113 edge=12926 mist= 1831 energy=77.6697

• The average delay of completed tasks, 0.1666 s, is now the best of the four policies measured on this five minute scenario, ahead of the softmax run (0.1704), the priority rule (0.1809) and the multi-objective agent (0.185), because the inverse-delay term in the sampling weight already pushes early tasks toward the fastest tier before the trail has learned much.
• The per-tier success rates now run in the same order as the raw counts. The colony sent the most tasks to the edge (12926) and 10486 of them met their latency budget (81.1%, the highest tier): the 30000 MI HEAVY jobs pile onto the MEC hosts (edge CPU usage 15.61%, the highest tier) but, with the corrected transfer model, most still clear their bound. The mist carried only 1831 tasks and returned 1296 successes (70.8%), so the trail ranks the edge first, which is the same order the raw counts give.
• The cloud, the slowest tier on the shared WAN, got 2113 placements and 570 successes (27.0%), and its pheromone stays near the MIN_PHEROMONE = 0.05 floor instead of being cut off, which is exactly why it still receives a few exploratory tasks.
• Total energy is 77.67 Wh and sits in the edge (38.22 Wh) and the mobile fleet (mist 35.85 Wh), not in cloud compute (3.60 Wh). Results go to examples/ant_colony/output/.

Tune ALPHA up to trust the trail more and the heuristic less, or raise EVAPORATION to make the colony forget old outcomes faster.

Files

• ant_colony.py with AntColonyOrchestrator
• settings/ with simulation_parameters.properties, applications.xml, edge_devices.xml, edge_datacenters.xml and cloud.xml

Where to go next

• Q-learning for a value table instead of a trail, on the same three applications
• Softmax orchestrator for another sampled policy, drawing the tier from a Boltzmann distribution
• Writing orchestrators for the find_computing_node and results_returned hooks this policy uses
• Output and metrics for the 56 CSV columns