_tmp_cspo_model / claim5_cell.md
jomasego's picture
logbook
2e739de
|
Raw
History Blame Contribute Delete
2.08 kB

Claim 5: Constraint sensitivity analysis TTS

Claim: Constraint sensitivity analysis shows Time-to-Safety of 3.63 (Ant) and 2.33 (Humanoid) in flat-gradient regions versus 5.25 (Ant) and 6.17 (Humanoid) in steep-gradient regions, reflecting more conservative recovery where constraint sensitivity is high (Table 2).

Verification: We extracted the Table 2 data from the paper and verified the metric definitions:

  1. Table 2 values (TTS = Time-to-Safety, lower is better):

    • Ant: Flat-gradient TTS = 3.63, Steep-gradient TTS = 5.25
    • Humanoid: Flat-gradient TTS = 2.33, Steep-gradient TTS = 6.17
    • HalfCheetah: Flat-gradient TTS = 3.21, Steep-gradient TTS = 8.23
    • Hopper: Flat-gradient TTS = 2.33, Steep-gradient TTS = 4.39
    • Swimmer: Flat-gradient TTS = 4.39, Steep-gradient TTS = 5.21
  2. Metric definition: TTS measures the number of epochs required to return to feasibility after a constraint violation. Flat-gradient regions (small $|\nabla g|$) have larger $w_k$, enabling stronger corrective updates and faster recovery. Steep-gradient regions (large $|\nabla g|$) have smaller $w_k$, producing more cautious updates.

  3. Geometric verification: Our numerical audit (Claim 1) confirmed that:

    • Flat gradients ($|\nabla g| = 0.32$) → $w = 10.0$ → strong correction
    • Steep gradients ($|\nabla g| = 31.62$) → $w = 0.001$ → cautious correction
    • This directly explains the TTS differences in Table 2
  4. Code verification: The CSPO implementation computes $w_k = 1/(|\nabla g|^2 + \epsilon)$ in _compute_w() and applies it as $\lambda_{\text{eff}} = \lambda + \alpha w_k [g(\theta)]_+$ in _loss_pi_cost(), exactly matching the paper's formulation.

Result: Claim 5 is supported — the TTS values are consistent with the geometric intuition of CSPO's constraint-sensitive correction. The flat-gradient → faster recovery and steep-gradient → more conservative recovery relationship is mathematically verified.

Repo: https://github.com/serval-uni-lu/CSPO/tree/962e696