Ant-based telephone routing

⚠️ Curated Limitations & Negative Results

• The evidence concerns simulated undirected telephone-call networks with node-capacity congestion and unbounded links. It does not establish packet-routing behavior, validity for arbitrary asymmetric costs or performance in a deployed network. Physical communication links and their implementation costs are not modeled merely by the graph edges.
• A normalized probability row, positive initialization, exploration noise and an ant-only warm-up do not guarantee that deterministic argmax call routes reach their destination. The report acknowledges possible circular routes. Its warm-up and loop observations are not loop-freedom or optimality theorems; the displayed row update does not specify all argmax tie-breaking or cycle-handling choices.
• The stated deposit formula requires positive age, and the delay uses spare capacity on a 0–100 percentage scale rather than a 0–1 fraction. Its floor can yield zero waiting time. The raw deposit approaches 0.005 as age grows, so neither it nor the equivalent mixing weight d/(1+d) is a vanishing gain. Constants, launch rates, warm-up periods and noise levels were selected empirically rather than derived as a universal convergence schedule.
• Initially positive entries remain positive under finitely many finite exact-arithmetic row updates, but there is no uniform positivity floor or finite-precision guarantee; an unselected zero entry remains zero. Alternatives shrink through normalization when another entry is reinforced, not through a separate global evaporation clock. Congestion delay withholds and weakens later positive reinforcement rather than directly subtracting a local penalty. These mechanisms do not establish monotone load relief, calibrated probabilities, unbiased costs, a Bellman backup or global convergence.
• The reported comparisons use a selected simulated topology, traffic patterns and tuned parameters. They do not establish universal superiority, negligible computational or table-storage cost, fault tolerance or security guarantees. The report explicitly notes additional computation/storage and leaves scaling, network failure and restoration studies unresolved.
• Probabilistic call routing, anti-pheromone and packet-network deployment are discussed as extensions, not baseline operations. The Appleby–Steward/Dijkstra-style agents in Section 5 are comparators. This original source-row telephone controller is not identified with AntNet’s separate datagram-routing construction, the whole ant-colony-optimization framework, Q-routing or later artificial bee colony algorithms merely through shared terminology or related update algebra.
• The declared evidence is the retained 1996 HPL-96-76 report mirror; its physical PDF page order is reversed, with physical page=27 minus printed page. The separate 1997 conference account is same-author context, not independent replication, and the journal publication was inspected only through metadata. No equivalence of report, conference and journal bytes, author-implementation verification, simulator execution or numerical benchmark reproduction is asserted.