Online Dynamic Cost-Driven Algorithm (On-Dyn-CDA)

• On-Dyn-CDA is a family of online cost-minimizing algorithms that sequentially update actions using aggregate feedback under time-varying and noisy conditions.
• The method leverages simple online linear regression to estimate key aggregate parameters, achieving fast convergence and logarithmic cumulative regret in demand response.
• Variants of On-Dyn-CDA have been tailored for vehicular task offloading and dynamic data replication, highlighting its scalability, efficiency, and privacy-preserving design.

Online Dynamic Cost-Driven Algorithm (On-Dyn-CDA) denotes an online decision rule in which actions are updated sequentially to minimize a cost objective under time variation and incomplete information. In the demand response formulation, On-Dyn-CDA is a joint online learning and pricing algorithm: a utility broadcasts a price signal, observes only aggregate noisy consumption response, and updates the next price by estimating aggregate parameters of users’ private quadratic cost functions, all with one round of communication per time slot (Li et al., 2017). The same acronym is also used in later work on vehicular task offloading and dynamic replication in distributed storage, where it names distinct online cost-driven policies adapted to those domains (Paknejad et al., 7 Sep 2025, Zuo et al., 28 Oct 2025). This suggests that On-Dyn-CDA is best understood as a domain-specific family of online, dynamic, cost-minimizing algorithms rather than as a single invariant update rule.

1. Demand response formulation

In the demand response setting, the operator faces time-varying demand response targets $d_t$, does not know the cost functions of consumers, and cannot have multiple rounds of information exchange with consumers (Li et al., 2017). At each time slot $t$, the operator sends a price signal $\lambda_t$ to users $i=1,\dots,N$, and each user responds by solving a local optimization problem under a private quadratic discomfort model

$$u_i(x_i^t)=\frac{1}{2}\beta_i(x_i^t)^2+\alpha_i x_i^t.$$

The observed response is noisy,

$\hat{x}_i^t=x_i^t+\epsilon_i^t,$

with $\epsilon_i^t$ modeled as Gaussian i.i.d. noise.

The operator seeks to minimize expected system cost over a finite horizon $T$, with a configurable normalized capacity $Y$. The objective used in the formulation is

$$\begin{aligned} \min_Y\min_{\bm{x}} \ &\sum_{t=1}^{T} \sum_{i=1}^{N} \frac{1}{N} \mathbb{E}\left[ \frac{1}{2} \beta_i (\hat{x}_i^{t})^2 + \alpha_i \hat{x}_i^{t} \right] \ &\quad + \sum_{t=1}^{T} \frac{1}{2N} \mathbb{E}\left[ \left( \sum_{i=1}^N \hat{x}_i^t - Y d_t \right)^2 \right] - \frac{\alpha Y T}{N}. \end{aligned}$$

The central modeling difficulty is that the operator knows neither $t$0 nor $t$1 a priori, while targets $t$2 vary over time and only one communication round is allowed per slot.

A defining feature of this formulation is that the operator does not attempt to identify all user-level parameters. Instead, the algorithm exploits the fact that the optimal pricing rule depends only on aggregate quantities. This aggregation is the main reduction that makes the online learning problem tractable in the stated model.

2. Learning-and-pricing mechanism

The demand response On-Dyn-CDA focuses on two aggregate unknowns,

$t$3

Under the aggregate observation model, the total observed response satisfies

$t$4

where $t$5 is summed i.i.d. Gaussian noise (Li et al., 2017). Because only $t$6 and $t$7 are required for price computation, the operator can estimate them from the history of aggregate observations $t$8 rather than from individual user measurements.

At each time step, the algorithm updates $t$9 by online linear regression of $\lambda_t$0 on $\lambda_t$1, using ordinary least squares and, for early time slots, possibly ridge regularization to avoid ill-conditioning. The next price is then computed by plug-in substitution,

$\lambda_t$2

Operationally, the algorithm cycle is: estimate aggregate parameters from prior slots, compute the current price from the estimated model and the target $\lambda_t$3, broadcast the price, collect the aggregate response $\lambda_t$4, and iterate.

Two properties of this mechanism are repeatedly emphasized in the formulation. First, only aggregate response is required, so scalability and privacy are retained. Second, the algorithm is explicitly online: the price at time $\lambda_t$5 is based only on information from slots $\lambda_t$6, while the response at slot $\lambda_t$7 becomes training data for later updates.

3. Regret analysis and asymptotic behavior

The performance criterion for the demand response On-Dyn-CDA is regret relative to an oracle that knows the true user parameters. The cumulative regret over $\lambda_t$8 periods is

$\lambda_t$9

and the per-slot expansion given in the analysis is

$i=1,\dots,N$0

with $i=1,\dots,N$1 and $i=1,\dots,N$2 problem-dependent constants (Li et al., 2017).

The main theoretical statements are that the per-slot gap decays as $i=1,\dots,N$3 and the cumulative regret satisfies

$i=1,\dots,N$4

The analysis attributes this behavior to consistency of the linear regression estimator: the estimator variance decays like $i=1,\dots,N$5, the price error variance and bias therefore decay as $i=1,\dots,N$6, and the system cost is quadratic in price error. The paper further states that logarithmic regret is optimal for this class of problems and cannot be improved in general.

This theoretical structure is significant because it separates estimation difficulty from control difficulty. The online policy is not directly optimized in a high-dimensional user space; instead, it is optimized through a low-dimensional aggregate linear model whose statistical error decays sufficiently fast to make the total regret only logarithmic.

4. Implementation and empirical behavior

The implementation burden of the demand response On-Dyn-CDA is deliberately small. The algorithm requires only simple least-squares regression with two parameters, aggregate consumption responses rather than individual measurements, and a single broadcast price per round (Li et al., 2017). For initialization, random or fixed initial prices can be used, and ridge regression is recommended in early rounds.

The reported simulations show fast learning dynamics. The online price quickly converges to the optimal price, typically within 50 rounds even with 100 users, and the difference between aggregate online response and optimal response quickly vanishes over time. The cumulative system regret closely follows a logarithmic trend, matching the theoretical prediction. The variance of price estimates dominates squared bias, and both decrease rapidly.

The empirical analysis also considers partial repetition in the target sequence $i=1,\dots,N$7. Even with repetitions in $i=1,\dots,N$8 up to 40%, the cumulative regret still grows sublinearly. This matters because repeated or slowly changing targets can make excitation of the regression problem weaker; the reported behavior indicates that the algorithm remains effective unless entire sequences are constant.

A common misconception in this setting is that an online price-learning scheme must recover each consumer’s individual preferences. The formulation explicitly avoids that requirement: only aggregate parameters are learned, and only aggregate response is observed. The resulting method is therefore presented as both tractable and scalable under the stated communication constraints.

5. Reuse of the name across domains

The acronym On-Dyn-CDA is reused in several later problem domains, but the underlying algorithms differ materially. The following summary collects the explicit usages appearing in the literature provided here (Li et al., 2017, Paknejad et al., 7 Sep 2025, Zuo et al., 28 Oct 2025).

Setting	Decision mechanism	Reported property
Demand response	Online linear regression of aggregate response, followed by plug-in price update	$i=1,\dots,N$9 cumulative regret
Vehicular networks	Greedy local-cost scheduling of arriving tasks to MEC servers	$$u_i(x_i^t)=\frac{1}{2}\beta_i(x_i^t)^2+\alpha_i x_i^t.$$0 s execution time in the most complex scenario; 3.42% lower task loss than Dynamic PSO; 29.22% lower average latency in complex scenarios
Distributed data access	Threshold-based creation, deletion, and migration of copies across servers	Competitive ratio $$u_i(x_i^t)=\frac{1}{2}\beta_i(x_i^t)^2+\alpha_i x_i^t.$$1

In vehicular networks, On-Dyn-CDA is a real-time task offloading algorithm for MEC-enabled vehicular systems (Paknejad et al., 7 Sep 2025). When a MEC server becomes available, the algorithm forms a decision window of arrived, not-yet-dropped tasks that could still meet their deadlines. For each candidate task, it simulates the assignment, evaluates the impact on other tasks in the window, and assigns the task with minimum local cost. The local cost penalizes forced drops and assignments that push tasks close to their maximum waiting time. The paper contrasts this design with static and dynamic PSO baselines, emphasizing that the offline static PSO benchmark excludes execution time and is infeasible for real-time applications despite its optimal performance in theory. It also states that the algorithm requires neither a dataset nor a training phase.

In distributed storage, On-Dyn-CDA is an online algorithm for dynamic replication over geo-distributed servers with heterogeneous storage costs (Zuo et al., 28 Oct 2025). After serving a local request at server $$u_i(x_i^t)=\frac{1}{2}\beta_i(x_i^t)^2+\alpha_i x_i^t.$$2, the algorithm keeps the copy for $$u_i(x_i^t)=\frac{1}{2}\beta_i(x_i^t)^2+\alpha_i x_i^t.$$3 time units; if another request arrives during that period, the timer is extended. When a copy would expire, it is deleted if other copies exist. If it is the only copy, the algorithm either keeps it indefinitely when $$u_i(x_i^t)=\frac{1}{2}\beta_i(x_i^t)^2+\alpha_i x_i^t.$$4, where $$u_i(x_i^t)=\frac{1}{2}\beta_i(x_i^t)^2+\alpha_i x_i^t.$$5 is the cheapest server, or immediately transfers it to $$u_i(x_i^t)=\frac{1}{2}\beta_i(x_i^t)^2+\alpha_i x_i^t.$$6, which then keeps it indefinitely. The paper proves a competitive ratio of

$$u_i(x_i^t)=\frac{1}{2}\beta_i(x_i^t)^2+\alpha_i x_i^t.$$7

where $$u_i(x_i^t)=\frac{1}{2}\beta_i(x_i^t)^2+\alpha_i x_i^t.$$8 is the max/min storage cost ratio among servers, and shows that no deterministic online algorithm can achieve a competitive ratio bounded by $$u_i(x_i^t)=\frac{1}{2}\beta_i(x_i^t)^2+\alpha_i x_i^t.$$9 for the general problem.

A recurrent source of confusion is therefore nomenclatural rather than technical: the same acronym denotes a learning-and-pricing algorithm in demand response, a greedy local-cost scheduler in vehicular MEC, and a threshold migration policy in distributed storage. This suggests that the stable content of the term is the online, dynamic, cost-driven viewpoint, not a single shared update equation.

6. Relation to adjacent online optimization frameworks

The design space surrounding On-Dyn-CDA overlaps with several broader online optimization frameworks. One nearby line of work is dynamic mirror descent, which incorporates a time-varying dynamical model $\hat{x}_i^t=x_i^t+\epsilon_i^t,$0 into online convex optimization and obtains regret bounds of the form

$\hat{x}_i^t=x_i^t+\epsilon_i^t,$1

for suitable conditions on the dynamics and step sizes (Hall et al., 2013). This framework addresses nonstationarity by measuring deviation from modeled dynamics rather than only total variation.

A second neighboring line is regularization-based online right-sizing of data centers. In heterogeneous data centers with operational cost and switching cost, an online regularization algorithm based on relative entropy solves a per-slot optimization problem and achieves a competitive ratio

$\hat{x}_i^t=x_i^t+\epsilon_i^t,$2

improving on the greedy ratio $\hat{x}_i^t=x_i^t+\epsilon_i^t,$3 (Zhang et al., 2018). With switching cost offset and linear operational cost, a specialized regularization algorithm obtains the competitive ratio

$\hat{x}_i^t=x_i^t+\epsilon_i^t,$4

Here the design emphasis is competitive analysis of switching-aware control rather than regret against a statistical oracle.

A third related direction is learning-based online resource allocation under unknown arrival counts. One-time and dynamic pricing algorithms learn dual prices from early arrivals or historical data and achieve $\hat{x}_i^t=x_i^t+\epsilon_i^t,$5-competitive performance under capacity conditions, without requiring prior knowledge of the total number of arrivals (Jaillet et al., 2012). That work explicitly contrasts its framework with earlier On-Dyn-CDA-style approaches that require knowledge of the number of arrivals under the random permutation model.

Taken together, these comparisons locate On-Dyn-CDA within a wider class of online cost-minimization methods that differ primarily in feedback structure and analysis criterion. Demand response On-Dyn-CDA is organized around aggregate regression and logarithmic regret; the data-center algorithm is organized around regularization and competitive ratio; dynamic mirror descent is organized around comparator tracking under a dynamical model; and learning-based online resource allocation is organized around dual-price estimation from arrivals. A plausible implication is that On-Dyn-CDA is most precisely characterized not by a specific optimization primitive, but by the conjunction of online feedback, dynamic system evolution, and explicit cost-driven decision updates.