Digital Pheromone Process Control

• Digital pheromone process control is a novel approach using two overlapping pheromone fields to coordinate agent behaviors in decentralized systems.
• It employs biologically plausible controllers such as spiking neural networks and finite-state machines to process sensory inputs and modulate actions.
• Empirical results demonstrate improved resource exploitation and faster response in both simulated and robotic swarm experiments through adaptive reaction–diffusion dynamics.

A double-pheromone mechanism with biologically plausible controllers is a swarm intelligence framework in which agents coordinate using two distinct environmental signaling fields (pheromones), managed by decentralized sensorimotor controllers that adhere to principles observed in biological systems. This architecture enhances both adaptability and robustness in collective behaviors such as foraging, aggregation, and escape, and has been demonstrated in computational models, physical robots, and hybrid experimental systems. Key research incorporates spiking neural networks (SNNs), finite-state machines, and explicit local sensory-motor loops to emulate the multi-timescale learning and stigmergic coordination of real social insects (Jimenez-Romero et al., 2015, Malíčková et al., 2015, Sun et al., 2019).

1. Biologically Plausible Agent Controllers

Biologically plausible controllers for double-pheromone systems fall into two main classes: spiking neural networks and compact sensorimotor state machines. In models such as "A Model for Foraging Ants, Controlled by Spiking Neural Networks and Double Pheromones," each agent is controlled by a SNN based on simplified leaky integrate-and-fire dynamics:

$$C_m \frac{du}{dt} = -g_L (u - u_{\text{rest}}) + \sum_j w_j\,\mathrm{psp}_j(t)$$

In this circuit, sensory input (olfactory, pain, food-contact) is transformed via STDP-trained pathways into motor outputs (forward, left, right). Additional neurons drive pheromone-release behaviors. Such SNNs support real-time associative learning and classical conditioning, consistent with insect neuroethology.

An alternative, exemplified by the off-lattice trail-following model in (Malíčková et al., 2015), uses a simple two-state finite-state machine to switch between "search-food" and "search-nest" modes. Sensory inputs are derived from paired antennae sampling local pheromone gradients and processed according to Weber's law. The deterministic turn-angle

$$\Delta\varphi = \begin{cases} +\pi/6 &\Gamma>\alpha \text{ and } c_R>c_L \ -\pi/6 &\Gamma>\alpha \text{ and } c_L>c_R \ 0 &\text{otherwise} \end{cases}$$

with $\Gamma = |(c_R-c_L)/(c_R+c_L)|$, modulates turning, while the amplitude of Gaussian noise in heading updates ($\sigma(c)$) is adapted according to cue strength.

In physical robot settings (e.g., ColCOS$\Phi$ (Sun et al., 2019)), local controllers consist of gradient-following logic, state transitions ("aggregate," "escape," "wander"), and proportional turning rules. State switching is triggered by sensory thresholds on the two pheromone channels.

2. Formal Structure of Double Pheromone Fields

Double-pheromone mechanisms maintain two overlapping, real-valued fields per spatial location, each serving distinct behavioral functions. The canonical example is the pairing of a positive-reinforcement pheromone (e.g., recruitment, aggregation, or food trail) with a negative-reinforcement pheromone (e.g., blocking, alarm, or avoidance).

Field evolution is typically governed by discrete-time, spatially distributed reaction–diffusion–decay equations:

$$\frac{\partial \Phi_p(x, y, t)}{\partial t} = -\frac{1}{e_p} \Phi_p(x, y, t) + d_p \Delta \Phi_p(x, y, t) + J_p(x, y, t)$$

where $\Phi_p$ is concentration, $e_p$ evaporation time constant, $d_p$ diffusion coefficient, $\Delta$ discrete Laplacian, and $$\Delta\varphi = \begin{cases} +\pi/6 &\Gamma>\alpha \text{ and } c_R>c_L \ -\pi/6 &\Gamma>\alpha \text{ and } c_L>c_R \ 0 &\text{otherwise} \end{cases}$$0 is injection from agent deposition. Local superposition supports arbitrary numbers of pheromone types in virtual testbeds (Sun et al., 2019), but most biological or robotic implementations focus on two.

Positive pheromone is generally associated with resource discovery or aggregation, deposited following a successful behavior (e.g., food contact (Jimenez-Romero et al., 2015), leader presence (Sun et al., 2019)). Negative pheromone is released under adverse conditions (e.g., failure to find food, detection of a predator), introducing a soft "blockade" through temporary repulsion or avoidance. Both fields decay exponentially but may differ in decay and diffusion rates to tune spatiotemporal influence.

3. Agent–Field and Field–Agent Coupling

Behavioral interaction between agents and pheromone fields is structured as a closed sensorimotor loop. Agents sense local concentration gradients (via mechanistic SNN readout, explicit antennas, or arrays of color sensors), compute steering and speed adjustments, update state machines, and deposit pheromones based on their internal states and external events.

The probability of choosing a path or segment is modulated by the positive and negative pheromone concentrations and, in some implementations, a heuristic:

$$\Delta\varphi = \begin{cases} +\pi/6 &\Gamma>\alpha \text{ and } c_R>c_L \ -\pi/6 &\Gamma>\alpha \text{ and } c_L>c_R \ 0 &\text{otherwise} \end{cases}$$1

Exponents $$\Delta\varphi = \begin{cases} +\pi/6 &\Gamma>\alpha \text{ and } c_R>c_L \ -\pi/6 &\Gamma>\alpha \text{ and } c_L>c_R \ 0 &\text{otherwise} \end{cases}$$2, $$\Delta\varphi = \begin{cases} +\pi/6 &\Gamma>\alpha \text{ and } c_R>c_L \ -\pi/6 &\Gamma>\alpha \text{ and } c_L>c_R \ 0 &\text{otherwise} \end{cases}$$3, and $$\Delta\varphi = \begin{cases} +\pi/6 &\Gamma>\alpha \text{ and } c_R>c_L \ -\pi/6 &\Gamma>\alpha \text{ and } c_L>c_R \ 0 &\text{otherwise} \end{cases}$$4 tune positive trail-following, heuristic exploitation, and negative trail avoidance, respectively (Jimenez-Romero et al., 2015).

Agent-based pheromone deposition is event-driven: e.g., food foragers lay positive pheromone when rewarded, negative pheromone when unsuccessful for some interval. Field decay ensures that old trails fade, enabling flexible re-exploration and adaptation to dynamic environments (Malíčková et al., 2015).

4. Empirical Performance and Parameter Sensitivity

A critical contribution of double-pheromone systems is improved task performance and adaptability over single-pheromone or positive-only ant colony optimization. In foraging experiments on a $$\Delta\varphi = \begin{cases} +\pi/6 &\Gamma>\alpha \text{ and } c_R>c_L \ -\pi/6 &\Gamma>\alpha \text{ and } c_L>c_R \ 0 &\text{otherwise} \end{cases}$$5 grid with 30 SNN-controlled ants and 20 food sources, double-pheromone agents achieved full resource depletion within 7,500 steps, compared to 90%+ food remaining for agents without pheromones, and a $$\Delta\varphi = \begin{cases} +\pi/6 &\Gamma>\alpha \text{ and } c_R>c_L \ -\pi/6 &\Gamma>\alpha \text{ and } c_L>c_R \ 0 &\text{otherwise} \end{cases}$$6 speedup in time-to-exhaustion over positive-only control (Jimenez-Romero et al., 2015).

In robotic implementations (e.g., swarm aggregation/escape in ColCOS$$\Delta\varphi = \begin{cases} +\pi/6 &\Gamma>\alpha \text{ and } c_R>c_L \ -\pi/6 &\Gamma>\alpha \text{ and } c_L>c_R \ 0 &\text{otherwise} \end{cases}$$7), group-level metrics such as mean Euclidean distance to a leader (for aggregation) or speed of dispersal (for escape) distinctively capture the temporal sequence and reestablishment of group organization under alternating pheromone dominance (Sun et al., 2019).

Parameter sweeps reveal critical relationships:

Key Parameter	Observed Effect	Optimal Regime
Negative pheromone decay ($$\Delta\varphi = \begin{cases} +\pi/6 &\Gamma>\alpha \text{ and } c_R>c_L \ -\pi/6 &\Gamma>\alpha \text{ and } c_L>c_R \ 0 &\text{otherwise} \end{cases}$$8)	Overseals if too slow; ineffective if too fast	$$\Delta\varphi = \begin{cases} +\pi/6 &\Gamma>\alpha \text{ and } c_R>c_L \ -\pi/6 &\Gamma>\alpha \text{ and } c_L>c_R \ 0 &\text{otherwise} \end{cases}$$9 (Jimenez-Romero et al., 2015)
$\Gamma = |(c_R-c_L)/(c_R+c_L)|$0 heuristic exponent	Maximal at $\Gamma = |(c_R-c_L)/(c_R+c_L)|$1; too high causes local overexploitation	$\Gamma = |(c_R-c_L)/(c_R+c_L)|$2
$\Gamma = |(c_R-c_L)/(c_R+c_L)|$3, $\Gamma = |(c_R-c_L)/(c_R+c_L)|$4 (diffusion/lifetime ratios)	$\Gamma = |(c_R-c_L)/(c_R+c_L)|$5, $\Gamma = |(c_R-c_L)/(c_R+c_L)|$6 promote agility and stable trail formation	$\Gamma = |(c_R-c_L)/(c_R+c_L)|$7\,mm$\Gamma = |(c_R-c_L)/(c_R+c_L)|$8/s, $\Gamma = |(c_R-c_L)/(c_R+c_L)|$9\,mm$\sigma(c)$0/s (Malíčková et al., 2015)
Noise-to-signal ratio ($\sigma(c)$1)	Governs exploration/exploitation balance	Signal-adaptive noise reduction

5. Collective Dynamics and Robustness

Double-pheromone architecture produces stigmergic memory supporting rapid formation and pruning of trails, clustering, and group escape. The combination of positive reinforcement and negative marking partitions the search space dynamically: rewarding paths are amplified, while unrewarding or dangerous regions are temporarily excluded but remain available for future revisiting because negative pheromones eventually evaporate (Jimenez-Romero et al., 2015, Malíčková et al., 2015).

Simulation experiments show trail formation, symmetry breaking among alternative food sources, large-scale synchronization of arrivals, and system adaptation to environmental changes (such as relocation or depletion of resources) on timescales matching pheromone decay (Malíčková et al., 2015). Robot experiments replicate aggregation-escape cycles by switching behavioral dominance between aggregation and alarm fields, confirming biological realism and robustness under noise, limited sensing, and simple controllers (Sun et al., 2019).

6. Biological and Engineering Plausibility

The double-pheromone framework is directly inspired by multi-cue communication in ants and other social insects, where multiple pheromones encode recruitment, blockades, alarm, or aggregation. Realism is ensured by: (a) local and noisy sensory processing, (b) decentralized one-step memory controllers, (c) minimal cognitive load and computation per agent, and (d) reaction–diffusion field dynamics with parameterizable timescales and spatial footprints (Jimenez-Romero et al., 2015, Malíčková et al., 2015, Sun et al., 2019).

In engineering contexts, the virtual pheromone substrates (e.g., 2D color screens) and minimalistic robot controllers enable physically realizable testbeds for scalable swarm experiments, with field implementation and controller modularity that generalize to more than two pheromones (Sun et al., 2019).

Such mechanisms yield emergent behaviors (trail formation, resource exploitation, group evacuation, regrouping) without central coordination, global state, or direct inter-agent signaling, matching the signature operational principles of collective animal behavior.

7. Limitations, Extensions, and Open Questions

Current double-pheromone models typically employ restricted sensory modalities (color-coded cues, light intensity, chemical gradients), limited plasticity rules (STDP), and static or quasi-static environments. A plausible implication is that integration of richer sensor arrays, neuromodulated or selective plasticity, and dynamical resource distribution would further advance both performance and biological verisimilitude.

The scalability to larger agent numbers, tuning of multi-pheromone interactions, and mapping to specific biological systems (e.g., different ant species' pheromone cocktails) remain active areas of research. The underlying formalism is also extendable to three or more signal channels, with corresponding growth in behavioral repertoire and task-switching capabilities (Sun et al., 2019).

Overall, double-pheromone mechanisms with biologically plausible controllers constitute a foundational paradigm for the design, analysis, and biological modeling of decentralized adaptive collective systems (Jimenez-Romero et al., 2015, Malíčková et al., 2015, Sun et al., 2019).