D2sac Asp Scheduling Eval

d2sac-asp-scheduling-eval

Diffusion-based Reinforcement Learning for Edge-enabled AI-Generated Content Services — Du et al. (2023) (arXiv:2303.13052, 2023)

What this evaluates

Evaluates a diffusion-based reinforcement learning agent's capability to dynamically assign AI-generated content tasks to edge service providers under stochastic workloads, optimizing for user utility while preventing system crashes and minimizing training time.

Datasets

• Custom AIGC Edge Simulation Environment — total 1000; splits: test (1000); repo https://github.com/Lizonghang/AGOD
• Gym Benchmark Tasks — total ?; splits: test (-1)

Metrics

• Cumulative Reward (primary) — range: other
  • Sum of step-wise rewards over an episode of 1,000 environment steps. Represents total user utility maximized by the scheduler.
• Finished Rate — range: percent
  • Percentage of successfully completed tasks out of the total 1,000 submitted tasks.
• Crashed Rate — range: percent
  • Percentage of tasks that failed or crashed during processing out of the total submitted tasks.
• Time to Baseline — range: hours
  • Training time (hours) required for the test reward to reach the reward level of the Crash Avoid heuristic policy.

Input / output format

Input: State vector containing ASP resource capacities, current task queue status, and network operation metrics.

Output: Discrete action selecting one of the 20 available ASPs to assign the incoming task.

Scoring recipe

def evaluate_episode(rewards, finished_count, crashed_count, total_tasks, baseline_reward):
    cumulative_reward = sum(rewards)
    finished_rate = (finished_count / total_tasks) * 100
    crashed_rate = (crashed_count / total_tasks) * 100
    time_to_baseline = next((i for i, r in enumerate(rewards) if sum(rewards[:i+1]) >= baseline_reward), None)
    return cumulative_reward, finished_rate, crashed_rate, time_to_baseline

Common pitfalls

• The exact mathematical formulation of the reward function is not provided, only described as maximizing user utility based on image quality assessment.
• Convergence metrics (Time/Step to Baseline) use the heuristic 'Crash Avoid' policy as an arbitrary reference, limiting cross-study comparability.
• Gym benchmark results for several baselines are cited from prior literature rather than re-evaluated under identical hyperparameters.

Evidence (verbatim from paper)

For the ASP selection problem, we summarize the best performance achieved by the proposed D2SAC and 11 benchmark policies in Table III, in terms of cumulative reward, training time, and convergence speed. ... The results presented in Table V offer a comprehensive comparison of several metrics, including finish rate, obtained utility, crash rate, and lost utility.

Citation

@misc{du2023diffusionrl,
  title={Diffusion-based Reinforcement Learning for Edge-enabled AI-Generated Content Services},
  author={Du et al. (2023)},
  year={2023},
  note={arXiv:2303.13052}
}

• arXiv: 2303.13052