Mec Offloading Eval

mec-offloading-eval

Lyapunov-guided Deep Reinforcement Learning for Stable Online Computation Offloading in Mobile-Edge Computing Networks — Bi et al. (2020) (arXiv:2010.01370, 2020)

What this evaluates

Evaluates the stability, convergence, and resource efficiency of online computation offloading algorithms in dynamic mobile-edge networks under stochastic task arrivals and time-varying channel conditions. It probes whether an algorithm can maintain queue stability and power constraints while maximizing computation throughput.

Datasets

• Simulated MEC Offloading Environment — total ?; splits: test (-1); repo https://github.com/revenol/LyDROO

Metrics

• weighted sum computation rate (primary) — range: other
  • Sum of per-device computation rates weighted by coefficients c_i: \sum_{i=1}^N c_i r_i^t. Measured in Mbps.
• average data queue length — range: other
  • Time-averaged backlog of task data at each wireless device, indicating stability. Stable if bounded over time.
• average power consumption — range: other
  • Time-averaged energy consumption per device, constrained by threshold \gamma_i.
• optimality ratio — range: [0, 1]
  • Ratio of LyDROO's per-frame objective value to the LyCD benchmark's objective value for subproblem (P2).

Input / output format

Input: State vector \xi^t containing channel gains, data queue backlogs, energy queue states, and task arrival rates for N wireless devices at each time frame t.

Output: Mixed-integer-continuous offloading action vector: binary offloading decisions x_i^t, transmission time \tau_i^t, local computation frequency f_i^t, and offloaded energy e_{i,O}^t for each device i.

Scoring recipe

# Initialize queues and accumulate metrics over T frames
for t in range(T):
    # Compute per-device computation rate r_i^t from action
    weighted_rate = sum(c_i * r_i^t for i in range(N))
    # Update data queue: Q_i^{t+1} = max(Q_i^t + A_i^t - r_i^t, 0)
    # Update energy queue: E_i^{t+1} = max(E_i^t + \gamma_i - e_i^t, 0)
    # Accumulate averages
avg_queue_len = mean(Q_i^t for t in range(T))
avg_power = mean(e_i^t for t in range(T))
optimality_ratio = objective_LyDROO / objective_LyCD

Common pitfalls

• Confusing short-term rate spikes with long-term stability; queue stability requires bounded backlog over thousands of frames, not just high instantaneous rate.
• Ignoring the power constraint violation; algorithms may achieve high rates but violate the average power threshold \gamma_i, which invalidates their feasibility.
• Misinterpreting the Lyapunov parameter V; larger V improves rate but increases queue length and power consumption, requiring careful tuning for practical deployment.

Evidence (verbatim from paper)

In Fig.5, we consider two data arrival rates with $\lambda_{i}=2.5$ and $3$ Mbps for all $i$, and plot the weighted sum computation rate, average data queue length, and average power consumption performance over time.

Citation

@misc{bi2020lydroo,
  title={Lyapunov-guided Deep Reinforcement Learning for Stable Online Computation Offloading in Mobile-Edge Computing Networks},
  author={Bi et al. (2020)},
  year={2020},
  note={arXiv:2010.01370}
}

• arXiv: 2010.01370