acceptodds
Withdrawn from ICLR 2027

Policy-Guided Monte-Carlo Tree Search for CAD-Grounded PCB Routing

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Learned printed circuit board (PCB) routing policies can produce suboptimal solutions on unseen boards. We propose PW-MCTS, a test-time search method that refines a pretrained routing policy through Monte Carlo tree search over native CAD operations. Board-wise iterated Bellman calibration aligns critic outputs with target-board returns. Engine determinization enables reproducible branch evaluation; incremental checkpointing, memoized branch-packed inference, and invalid-action filtering reduce search overhead. We evaluate PW-MCTS on a subset of PCBWorld-Bench, an open-source PCB dataset, using a policy trained only on synthetic boards. At equal draws, PW-MCTS raises the fraction of attempts that complete a board without error-severity design-rule violations from 42.2% to 63.2%, and mean routability from to . Under matched rollout-time budgets after calibration, search raises board coverage from 84.7% to 89.3% with eight simulations per decision and roughly one-quarter as many trajectories as policy sampling. These results support test-time refinement by balancing search effort and repeated sampling, with attainable performance still shaped by the routing MDP and prior quality.

Primary area:
reinforcement learning
Keywords:
PCB routing, MCTS, CAD-grounded routing, policy improvement, value calibration

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_QqCbOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes PW-MCTS, a test-time search method that refines a pretrained PPO routing policy through Monte Carlo tree search over native CAD operations in PCBWorld. The method uses board-wise iterated Bellman calibration to align critic outputs with target-board returns, determinizes the CAD engine for reproducible branch evaluation, and introduces incremental checkpointing, memoized branch-packed inference, and invalid-action filtering to reduce search overhead. Experiments are conducted on 50 open-source PCBs from PCBWorld-Bench, using a policy trained only on synthetic boards. The paper reports improvements over the same pretrained PPO policy under equal draws and matched rollout-time budgets.

Critical strengths
  1. Test-time refinement over native CAD operations is a reasonable direction. Instead of retraining the policy or searching in a grid abstraction, PW-MCTS searches directly over engine-level CAD actions. This is a sensible way to study test-time policy improvement for PCB routing. However, it is open to question whether doing so guarantees a legal outcome.
  2. The engineering optimizations are useful. Determinizing the CAD engine, incremental checkpointing, memoized branch-packed inference, and invalid-action filtering address real practical bottlenecks in CAD-grounded search. These contributions could be valuable for future work in this area.
Critical weaknesses
  1. The method does not adequately address practical PCB design constraints, and the reported layouts appear to violate basic manufacturability requirements. The paper relies on engine-level DRC, action masking, and invalid-action filtering, but it does not provide a mechanism for users to manually add or prioritize custom design constraints, such as angle restrictions, differential-pair rules, impedance requirements, length matching, keepouts, or layer-specific constraints. More importantly, from the rendered layouts, I observe what appear to be 90-degree traces (e.g., board 0344) and potential shorts (e.g., board 0122). If confirmed, such layouts would be directly unusable in practice. The paper should clarify whether these layouts pass all relevant DRC checks and whether “clean pass” truly covers all error-severity violations. As presented, the methodology cannot guarantee that the final layout satisfies all constraints or flexible user requirements. Even without enforcing stricter constraints, PW-MCTS is only comparable to or worse than Freerouting. If such constraints were enforced, performance would likely degrade further.

  2. The experimental comparison is not convincing, and the definition of routability metric is wrong. The definition of “routability” is confusing. Actually, the definition of Eq. 6 cannot be described as "routability".In addition, the only classical external baseline is Freerouting, which is a single A*-based tool and may not represent stronger classical routers. For example, the incomplete route in case 0134 may be an artifact of Freerouting’s limited parameterization rather than a fundamental limitation of classical routing. Stronger baselines such as pcbrouter, tuned KiCad routers, or commercial routers should be compared, ideally under matched time or computation budgets. Table 11 already shows that PW-MCTS is below Freerouting on several metrics; this undermines the claim that the proposed method is a competitive routing approach.

  3. The convergence and stability of the Transformer-based policy are central but insufficiently explained. Given the heterogeneous token types and the factored action space, the paper’s key premise is that the model can converge and produce stable actions. However, the paper mainly lists architecture and PPO hyperparameters. It lacks training curves, policy entropy, gradient norms, value-loss behavior, seed variance, action-distribution stability, and failure-mode analysis. Without these diagnostics, it is difficult to assess whether the model genuinely converges in a stable way or whether the reported behavior is fragile.

Questions for authors
  1. For Figure 11, especially boards 0122 and 0344, please provide DRC reports. Do these layouts contain 90-degree traces or shorts? If so, how do they pass the paper’s clean-pass criterion?
  2. How can a user add or enforce custom design constraints, such as angle limits, impedance, differential pairs, length matching, keepouts, or layer restrictions? What happens to performance when such constraints are enforced?
  3. Why does PW-MCTS underperform Freerouting in Table 11 even without strict constraints? Can the authors provide matched-time comparisons against Freerouting and other classical routers such as pcbrouter?
  4. What evidence can the authors provide for convergence and stable action output? Specifically, training/validation curves, policy entropy, gradient norms, seed variance, and action-distribution stability across checkpoints would help.
Reviewer_vu8cOpenReview ↗
Rating
3
Confidence
1
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes PW-MCTS, a test-time search method that refines a pretrained routing policy through MCTS over native CAD operations.

Critical strengths

The consider problem is practical and meaningful. The idea of using MCTS really makes sense.

Critical weaknesses
  • The method is more like a combination of some existing methods, not that noval.
  • The method is not compared with commercial tools.
  • The improvement is not that significant, and the additional end-to-end time cost is unclear.
Questions for authors

See weaknesses.

Reviewer_CFp6OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper applies Monte Carlo tree search (MCTS) to the PCB routing problem. It works with a pre-trained routing policy obtained from PPO training. It first applies an iterative Bellman to calibrate the pre-trained critic. The paper documents many optimization for the tree search to ensure reproducible and efficient search, such as reproducing routing states across search branches. Empirical evaluations on 50 open-source PCB routing tasks show improvement over the policy rollout baseline.

Critical strengths

Applying MCTS to PCB routing is an interesting application. Combining search and a policy is proven to be effective in addressing large search space so the application to PCB routing is natural. Empirical evaluations validate this idea: under both equal-draw and matched-time evaluations show improvements over the baseline.

Critical weaknesses
  1. There is only one baseline based on PPO policy rollout. As this paper proposes a search algorithm improvement, we should expect at minimum comparing with other standard search algorithms, such as best-first search and beam search. It is crucial to be comprehensive in the baseline selections to properly evaluate the effectiveness of the proposed MCTS approach.

  2. The paper is difficult to understand at several places. Please see the list of clarification questions below. The main paper makes lots of references to sections in the appendix. That seems to show a better organization is necessary in order to put most relevant information in the main component.

Questions for authors

Clarification questions:

  1. Section 3.1: what is the difference between DRVDRV and DRVerrDRV_{err}?

  2. The definition of D(s)D(s) in appendix B refers to a parameter o>0o>0 that sets the knee. It needs more information on what this means.

  3. Line 207: g^(Vdone)\hat{g}(V_{done}), what is the rationale between using the median calibrated value?

  4. Section 5.1: the baseline. I am confused by the description: first PPO rollout on the same 16 rollout seeds. Is it rolling out the pre-trained policy network 16 times with fixed seeds?

  5. In the following sentence on matched time: what is a trajectory? Does it refer to a complete MCTS tree?

  6. I am also confused by the iterative Bellman calibration step. The motivation is to transform the critic estimate into the same measurement used by the MCTS. However, the calibration seems to involve quantities used in PPO training only? Unless rir_i does not refer to PPO training rewards?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Discussion (0)

Sign in to comment.