Continual Learning for Reliable In-Sequence Molecular Communication

I. INTRODUCTION

• Simulation studies supplement analytical work as communication settings become more complex, and machine learning can estimate measures such as latency and error rate from prior simulations.Using ML estimates can reduce the computational cost of additional simulations.Source: §I · p. 1
• Sequentially introducing more complex settings requires updating models with new simulation results without losing performance estimates for earlier settings.The paper illustrates this challenge as forgetting when training progresses from 1D to 2D and then 3D settings.Source: §I · p. 1
• The method is evaluated for latency estimation in a feedback-based protocol that sends and receives information molecules in a specified order.The protocol uses SW-ARQ, described as feasible with simple computation and memory functions in bio-nanomachines.Source: §I · p. 1
• AMR stores past simulation samples, synthesizes additional samples, and replays both to regularize updates as experimental settings change.This approach is intended to retain earlier knowledge while adapting to new estimation tasks.Source: §I · pp. 1–2
• AMR maintains higher overall RTT estimation accuracy than existing continual-learning methods across sequentially learned estimation tasks.The reported gains reduce forgetting while preserving competitive learning plasticity under moderate training cost.Source: §I · p. 2

II. BACKGROUND AND RELATED WORK

• Reliable in-sequence delivery can support artificial morphogen control in tissue growth and control the order and duration of drug-molecule exposure.The paper connects these applications to guiding bio-nanachine growth and modulating tumor behavior.Source: §II · p. 2
• Earlier feedback-based rate-control and molecule-delivery protocols use SW-ARQ, but the cited protocols do not use machine learning.The paper positions its work as extending this line of protocol research with continual-learning performance estimation.Source: §II · p. 2
• Prior molecular-communication ML studies address regression and classification tasks but do not employ continual learning.Their models require retraining from scratch and may forget when incrementally trained across experimental settings.Source: §II · p. 2
• Unlike the authors’ prior study of existing CL methods in molecular communication, this work develops a new method and evaluates it against existing methods..Source: §II · p. 2
• AMR combines rehearsal of past and synthetic samples with replay-loss regularization to discourage parameter updates that degrade replay performance.This is a hybrid of the replay and regularization approaches in the paper's CL taxonomy.Source: §II · p. 2

III. MOLECULAR COMMUNICATION MODEL

• The communication setting is a bounded 3D aqueous medium in which Tx and Rx exchange information and acknowledgement molecules, with DNA used as a possible carrier motivated by wet-lab experiments.Source: §III · p. 2
• The model includes diffusive, directional, and hybrid transport; in the hybrid case, information molecules travel directionally while ACK molecules diffuse.Source: §III · pp. 2–3, Figure 1 · p. 3
• Diffusion is modeled as random thermal motion governed by the diffusion coefficient D, while directional transport uses a microtubule to guide molecular motors.Molecules attached to motors travel at a constant velocity for an expected distance before diffusing away.Source: §III · p. 2
• Bio-nanomachines are modeled as modified cells with transmission, reception, logic, memory, and timing functions for molecule handling and programmed responses.Source: §III · p. 3
• Under SW-ARQ, the Tx sends duplicated information molecules and advances to the next message after receiving an ACK; timeouts trigger retransmissions.The Rx can retransmit ACK molecules for the previous message if it receives no information molecule for the next message within the timeout interval.Source: §III · p. 3

IV. CONTINUAL PERFORMANCE ESTIMATION

• AMR trains an RTT estimator on a stream of simulation datasets, with experimental settings as inputs and RTT as the estimation target.Its objective is to minimize cumulative training loss over the tasks observed so far.Source: §IV · p. 3
• AMR combines current-task loss with replay losses from past and synthetic samples to regularize parameter updates against degraded replay performance.The replay losses can also allow current-task updates to improve performance on earlier tasks; α and β control the influence of the two replay buffers.Source: §IV · p. 4
• AMR uses reservoir sampling to keep the primary buffer bounded while maintaining a representative set of past samples.A new sample is inserted with probability C/n and, if selected, replaces a uniformly chosen stored sample.Source: §IV · p. 4
• AMR selects interpolation or extrapolation using a random factor: values from 0 to 1 select interpolation, and other values select extrapolation.The factor is sampled over a range controlled by q.Source: §IV · p. 4
• Interpolation creates synthetic inputs between two past inputs and assigns targets by distance-weighted interpolation, broadening replay coverage within their convex neighborhood.Source: §IV · p. 4
• Extrapolation creates synthetic samples beyond the segment between two past inputs while extending the directional trend between their input-target pairs.Source: §IV · p. 4

• AMR, CLeaR, ACLT, and A2ER are evaluated through Monte Carlo particle-based molecular-communication simulations.Source: §V · p. 5
• AMR is implemented as a two-layer feedforward neural network with a fully connected ReLU hidden layer, and inputs and outputs are min-max normalized to [0, 1].Source: §V · p. 5
• AMR and CLeaR hyperparameters are selected by grid search on a training-data subset, while ACLT and A2ER settings come from their original papers,.Source: §V · p. 5
• All methods use stochastic gradient descent for 100 epochs, batch size 1, and learning rate 0.001, with mean squared error and early stopping.Early stopping uses patience 10 and minimum improvement threshold 0.003.Source: §V · p. 5
• Table I presents the hyperparameter configuration for AMR and the comparison methods.Source: Table I · p. 5, §V · p. 5

• Each training and test sample uses the median RTT from 500 independent simulations.Source: A. Simulation Settings and Evaluation… · p. 5
• The simulation treats boundaries and molecule-to-molecule collisions as non-repulsive, while preventing molecules from occupying another molecule’s location.Source: A. Simulation Settings and Evaluation… · p. 5
• Models are trained incrementally on 12 estimation tasks, with settings and RTT targets prepared from simulated molecular transport.The task sequence is shown in Fig. 3; training inputs combine Tx-Rx distance, noise level, and other experimental settings.Source: A. Simulation Settings and Evaluation… · p. 5, Figure 3 · p. 5
• Table II identifies the simulation settings used for molecular communication.Source: Table II · p. 5
• During training, each method accesses only the current task’s data, but at test time estimates RTT for the current and all prior tasks without task labels.Source: A. Simulation Settings and Evaluation… · p. 5
• The 12 evaluation scenarios assess each method after sequential training on t tasks, using mean squared error between estimated and true RTT.Matrix rows are scenarios and columns are tasks; an entry can report a task’s error after training on all tasks.Source: A. Simulation Settings and Evaluation… · p. 5
• Lplasticity averages diagonal test errors to measure rapid adaptation, while Lstability averages errors across all tasks after the final training scenario to measure retention.Lower values of both metrics indicate better estimation accuracy.Source: A. Simulation Settings and Evaluation… · p. 6
• F measures average task-wise test-error change from initial learning to the end of training, whereas Fr measures the average relative increase.Higher F indicates greater forgetting; Fr = 1 means final error is twice initial error, and Fr = 0 means no forgetting.Source: A. Simulation Settings and Evaluation… · p. 6

B. Results

• AMR achieves the lowest Lstability and F among four methods, with statistically significant Lstability gains under both tests (p < 0.05).AMR also has the lowest standard deviation for Lstability and F.Source: Table III · p. 6, B. Results · p. 6
• A2ER has the lowest Lplasticity, while AMR is second-lowest with the lowest standard deviation; their difference is small.AMR therefore maintains competitive plasticity while improving stability.Source: B. Results · p. 6
• AMR has the lowest average forgetting ratio, 0.1384, versus 0.1518 for ACLT, 0.1596 for A2ER, and 0.1662 for CLeaR.Its average rank across scenarios is 1.9, compared with 2.1, 3.0, and 3.0, respectively.Source: Figure 4 · p. 6, B. Results · p. 6
• AMR's mean Lstability is lowest at moderate training cost, while its total training time is comparable to CLeaR and A2ER and below ACLT.CLeaR has the lowest training time, followed by A2ER and AMR.Source: B. Results · p. 6, Table IV · p. 6, Figure 5 · p. 7
• Across tested q settings, Lstability is similar; q = 0.1 gives the lowest Lstability, q = 0 the lowest Lplasticity, and q = 0.2 the lowest F.The forgetting curves are broadly similar, with moderate q values tending to mitigate forgetting favorably.Source: B. Results · pp. 6–7, Table V · p. 7
• For q values, total training times range from 8.76 to 10.29 seconds; q = 0.1 gives the lowest time and best reported cost-performance tradeoff.The reported q = 0.1 configuration also has the lowest Lstability while maintaining low F.Source: B. Results · p. 7, Table VI · p. 7

VI. CONCLUSION

• AMR significantly outperforms existing CL methods in maintaining overall estimation accuracy across sequentially learned estimation tasks.Source: §VI · p. 7
• AMR synthesizes replay samples through interpolation and extrapolation, stores past and synthetic samples in bounded buffers, and uses replay losses to regularize model updates.Source: §VI · p. 7
• Future work will extend AMR to mobile bio-nanomachines, time-varying environments, and larger task sequences, and investigate adaptive q selection and broader replay-sample generation.Source: §VI · p. 7