acceptodds
Under review as a conference paper at ICLR 2027

Decision-TabPFN: Separating Action Scoring from Inventory Control

Abstract

A frozen predictor can select low-loss actions on fixed queries yet yield a weaker sequential policy. We study this gap in perishable inventory using Decision-TabPFN, which scores immediate profit from completed transitions. Matched comparisons show a TabPFN advantage over XGBoost on Decision-TabPFN histories, but this advantage varies with the source of queried states. Holding the backbone fixed, independently selected forecast-to-order pipelines earn more profit in two of three regimes. Fixed-budget context comparisons show that offline rankings need not persist online. To connect prediction with control, we bound inventory-dependent continuation value independently of the remaining horizon and cancel forecasting error shared by policies. The resulting inequality separates demand-prediction error from model-planning error and yields sublinear dynamic-optimum regret under explicit cyclic-demand learning and complete-horizon planning. A separate block-planning study exposes strong prior sensitivity and no demonstrated benefit from TabPFN proposals. Together, the results distinguish action scores on deployment states, the supplied decision structure, and continuation value as separate evaluation targets.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Tabular foundation models, In-context learning, Sequential decision-making, Inventory control, Regret analysis

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_8qe6OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates the gap between predictive accuracy and sequential decision-making in perishable inventory control using Decision-TabPFN. Through controlled experiments, the authors analyze predictor quality, decision interfaces, and context selection strategies. The paper further derives an inventory-specific value bound and establishes a regret guarantee that separates demand-prediction error from planning error.

Overall, the paper is primarily an empirical analysis of existing challenges, without providing a convincing solution to address the identified limitations. Moreover, the empirical and theoretical contributions are not sufficiently connected, making the overall contribution of the paper less coherent.

Critical strengths
  1. The experiments cover multiple inventory settings, decision interfaces, and context selection strategies.

  2. The theoretical analysis provides an interesting perspective on how prediction and planning errors affect long-term policy performance.

Critical weaknesses
  1. Limited methodological novelty. The first two contributions primarily provide empirical analyses of the discrepancies between predictive accuracy, decision-interface performance, and online policy returns. However, these issues have been extensively studied in sequential decision-making. The paper does not sufficiently clarify what new insights its empirical findings provide beyond existing knowledge, particularly since the investigated alternatives largely rely on established decision-making principles.

  2. Limited theoretical novelty and practical relevance. As acknowledged in the paper, Gong and Simchi-Levi [1] have already established square-root-order regret guarantees for inventory control with cyclic demand. Although this paper derives an inventory-specific value bound and a different regret analysis, the significance of these theoretical advances over existing results is not sufficiently established. Moreover, the guarantee relies on an explicit demand model and complete-horizon planning, whose computational scalability remains limited. This raises concerns about the novelty and practical applicability of the theoretical contribution.

  3. Weak connections among the contributions. The empirical findings and theoretical analysis are not sufficiently integrated. In particular, the regret guarantee is established for a separate planning extension and does not rely on any TabPFN-specific properties. Furthermore, the experiments do not demonstrate a clear advantage from using TabPFN proposals within the planning extension. Consequently, the relevance of the theoretical analysis to the proposed Decision-TabPFN framework remains unclear.

Reference

[1] Gong, X.-Y., and Simchi-Levi, D. Bandits atop Reinforcement Learning: Tackling Online Inventory Models with Cyclic Demands. Management Science, 70(9): 6139–6157, 2024.

Questions for authors
  1. In the AC–FO comparison, both the prediction target and decision rule change. Can the authors provide further ablations to isolate the contribution of each component?

  2. For coverage-based context selection, can the authors separately evaluate state-similarity retrieval and action balancing?

  3. What is the computational overhead of the planning extension at larger action spaces? Can the authors clarify whether TabPFN proposals provide any measurable efficiency advantage over simpler alternatives?

2 comments
AuthorsOpenReview ↗

What inventory history tells a frozen scorer, and what one-day scoring misses

Thank you for pressing us on what is specific to perishable inventory and how the theory bears on Decision-TabPFN. Our direct action-scoring policy (AC) uses frozen TabPFN to predict the profit of candidate orders from operating records. The revision studies what those records contribute, then measures what even an accurate one-day profit score can leave out.

W1. What did we learn beyond the distinction between prediction and control?

For a perishable product, two inventories of the same size can have different amounts left to sell tomorrow. Under backlogging, newest-first issuance and a fixed order fee of 20, AC earned 515.74 per day using the full operating record, 462.11 when the age-specific inventory features were removed and 347.76 when only orders and profits were retained. All three figures average days 1 through 10,000, including initialization (Table 12).

The timing of past records also mattered. Across the ten positions in the hidden demand cycle, the correlation between average order size and average demand in the final third of the runs ranged from 0.88 to 0.98 for AC, compared with −0.97 to −0.92 for the conventional stock rules (Figure 6). At fixed inventories under lost sales and oldest-first issuance, removing past records from the same cycle position increased the chosen order's expected one-day profit loss by 27.6%, compared with 3.9% after removing the same number at random (Appendix G.3). Only the evaluator saw cycle labels.

In lost sales, AC beat median forecast-to-order (FO) in all four settings. Neither cost-optimizing FO contrast passed Holm correction (Tables 1, 3, 15).

W2. What does finite shelf life add to existing regret results?

Gong and Simchi-Levi (2024) give a square-root-order regret guarantee for cyclic nonperishable inventory with a known period and discuss unknown length with a known upper bound in Remark 2. Our planning theorem also assumes a known upper bound. The new step is to control how the stock left by today's order affects future value when units have different expiration dates. Under lost sales, Lemma 6.2 bounds the difference in optimal remaining profit between two starting inventories after the same demand history, independently of the days remaining. Without expiration or disposal, a single excess unit under zero demand can incur a holding charge every remaining day (Appendix A.13).

Past demand remains informative after the original inventory expires. Subtracting the value of empty stock after the same demand history lets the history-only contribution cancel in expectation between policies, because orders do not affect demand (Appendix A.5). This gives prediction error a coefficient independent of the remaining horizon. Theorem 6.3 uses that bound to separate demand estimation from the error allowed in full-horizon planning when demand is observed or recoverable.

W3. Which results apply to the deployed scorer?

At inventories visited by AC, Proposition A.3 bounds immediate scoring and sampling losses and the effect of the resulting inventory on future value, without requiring a gap between orders. Table 14 measures the corresponding economic quantities in eight small lost-sales instances. With three-day shelf life, oldest-first issuance and a fee of 20, AC's executed order gave up 1.866 in immediate profit per day relative to the best one-day order. Even that best one-day order gave up 3.752 relative to the full-horizon choice when both were followed by optimal orders. The table also reports the signed difference in continuation value between AC's order and the best one-day order, which is needed to recover AC's total loss. In both three-day instances with this fee the myopic loss had the larger mean. The immediate loss had the larger mean in the other six.

Corollary A.4 gives sufficient conditions for sublinear AC regret. It requires a uniform positive gap between the two best immediate orders and sublinear cumulative contributions from score error, sampling and the future value forgone by the best immediate order. For the default ten retained orders, let sts_t be the standard deviation of their scores and τt\tau_t the sampling temperature. The effective temperature in profit units is λt=τt(st+10−6)\lambda_t=\tau_t(s_t+10^{-6}). Since τt≥0.1\tau_t\geq 0.1, the proof's daily sampling upper bound satisfies λtlog⁡10≥10−7log⁡10\lambda_t\log 10\geq 10^{-7}\log 10 after initialization. Summing that bound cannot certify sublinear regret for the default sampler. Actual loss may be smaller. Proposition A.5 gives an alternative certificate using valid bounds on score error and on the probability that two orders leave different inventories, without computing the optimal continuation. The deployed checkpoint has not been shown to satisfy these sufficient conditions. Theorem 6.3 gives a separate guarantee for the planner that checks each proposal against its full-horizon alternative.

Q1. How far do the comparisons separate the prediction target from the ordering rule?

When AC also selects greedily, AC and cost-optimizing FO apply the same decision rule: each picks the order with the highest estimated expected profit for the coming day. They differ only in what the frozen TabPFN predicts, either the realized profit of a candidate order (AC) or next-day demand, which the known cost formula converts into expected profit over 129 forecast quantiles (FO). The right-hand column below therefore compares the two prediction targets under a common decision rule. Entries are mean daily profit differences on ten paired 10,000-day demand paths per setting (Table 3).

Inventory settingTop-ten AC minus FOGreedy AC minus FO
Lost sales, oldest first, no fee+7.65±20.92+7.65 \pm 20.92+13.66±18.70+13.66 \pm 18.70
Lost sales, newest first, no fee−10.72±10.74-10.72 \pm 10.74+1.01±4.56+1.01 \pm 4.56
Backlogging, newest first, fee of 20−21.54±4.43-21.54 \pm 4.43−17.98±3.40-17.98 \pm 3.40

Margins are paired 95% confidence interval half-widths. Only the backlogging difference passed Holm correction, in both columns. The sampler itself mattered little: greedy minus top-ten AC was +6.01±31.79+6.01 \pm 31.79, +11.73±9.66+11.73 \pm 9.66 and +3.56±4.83+3.56 \pm 4.83 on these paths, none significant after Holm correction, and on the Table 1 paths greedy selection earned slightly less than the default (Table 12). In the backlogging setting the FO advantage therefore comes from predicting demand and applying the cost formula. These two cannot be separated further, because a demand forecast needs some rule, such as the cost formula or an order-up-to target, to become an order. Table 11 shows how much that rule matters: changing the safety-stock multiplier from 0.2 to 1.0 reversed the ranking against AC in both settings tested.

Q2. Which component of coverage-based retrieval helped?

We withdraw the claim that coverage-based retrieval improves online profit over uniform sampling in the two newest-first settings. Selecting similar inventories and balancing historical order sizes changed together, so the original comparison could not assign the gain to either component. The deployed AC uses every completed record. Table 12 now addresses the practical question for that policy by varying memory length and recorded features directly.

Q3. How much does planning cost, and do TabPFN proposals speed it up?

Direct scoring. At the largest grid tested, scoring roughly 800 candidate orders took about three seconds per decision (Figure 4).

Planning cost. For shelf life mm and maximum order AA, the capped demand has d=mA+1d = mA + 1 possible values. With CC demand-mixture components, exact two-day planning takes O(Cd2+Ad)O(Cd^2 + Ad) operations and O(d2)O(d^2) working memory beyond the period counts (Appendix A.8). For fixed mm and CC, this gives a quadratic upper bound in AA.

For a horizon of TT days and a period upper bound Hmax⁡H_{\max}, exact full-horizon planning requires at most T(A+1)m(T+d)MT(A+1)^m (T+d)^M state updates, where M=(d−1)Hmax⁡(Hmax⁡+1)/2M = (d-1)H_{\max}(H_{\max}+1)/2 (Appendix A.12). On the main grid, m=3m = 3, A=600A = 600 and Hmax⁡=10H_{\max} = 10, giving d=1,801d = 1{,}801 and M=99,000M = 99{,}000. This count is prohibitive at that scale, so on the main grid the full-horizon theorem is a statistical guarantee rather than a practical planner (Appendix A.6). We ran exact full-horizon checks on small instances with binary orders over 15 or 31 days.

Do TabPFN proposals help? Two-day planning shares its conditional demand calculations across orders. A TabPFN proposal adds an inference call and gave no detectable profit gain over ranking orders by expected one-day profit. In the full-horizon checks, the two rules chose identical orders throughout 79 of 80 paired runs and differed on one day within tolerance in the remaining run (Appendix F.1). Whether a proposal saves enough search time to offset its inference cost on a large grid remains unmeasured.

Reviewer_HdtSOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper examines the use of a frozen predictor based on TabPFN, a pretrained transformer based model that relies on in-context learning. as 'Decision-TabPFN' that treats inventory decisions as a regression problem, through 'imitation learning'. This is attractive because DecisionTabPFN requires no policy training, demand modeling or inventory formulas such basestock, EOQ etc. Prior approaches to such inventory planning problems were the classical predict-then-optimize where we first predict demand and then, for example, embed this within a (stochastic) inventory optimization model to determine the best inventory decisions. RL is another approach.

This is certainly an important an interesting problem to analyze and has immense practical value. if Decision-TabPFN matches or outperforms (or even comes close to) prior approaches that clearly require a lot more effort and resources, that would be a valuable result.

The paper setups up several experiments to systematically decomposing the problem structure to answer questions comparing the performance of Decision-TabPFN to baselines. They also prove an elegant bound result that is useful in the context of perishable inventory planning, which is the focus of this paper, and is a valuable contribution in itself.

Critical strengths

The problem being addressed, the experimental rigor, and the bound result are impressive. The honesty in presenting negative/neutral results without trying to cherry-pick is another big positive and a valuable result.

Critical weaknesses

Decision-TabPFN does not appear to exhibit a strong performance, and the work ends up looking more like a diagnostic paper on why this does not do well in different settings. There is no clear recommendation on the way forward- e.g., what Decision-TabPFN maybe in some hybrid combination with prior classical approaches can positively improve over prior approaches.

Theoretical result is disconnected from the practical experiments. Several sub-studies that do not come together and lead us to any decisive outcome. That the performance of Decision-TabPFN is underwhelming even for a basic inventory planning problem with a limited problem scope raises generalizability and scalability concerns.

Questions for authors
  1. Have you tried a simple predict-then-myopically optimize with TabPFN as the demand forecast generator rather than directly as a decision generator?
  2. The sparse-prior myopic controller appears very competitive. Why is this not a primary baseline, and also does its strong performance suggest that the underlying forecast problem is easy?
  3. are there practical scenarios where Decision-TabPFN as implemented will be preferable over predict-then-optimizer with the same backbone - real world practical settings where it can statistically dominate prior approaches?
2 comments
AuthorsOpenReview ↗

When direct scoring and forecast-based ordering help in perishable inventory

Thank you for recognizing the inventory-value bound and for asking when a retailer would choose Decision-TabPFN. Our direct action-scoring policy (AC) uses realized profits from the retailer's own orders as regression labels. It requires no expert demonstrations and keeps TabPFN's pretrained weights fixed. Forecast-to-order (FO) uses the same frozen model to predict demand, then converts the forecast into an order.

W1. How competitive is AC, and when would a hybrid approach help?

Over 10,000 days, AC earned higher mean daily profit than Thompson sampling, XGBoost forecasting and the noncyclic stock rule in all eight settings. Table 1 gives the paired comparisons below. The last column uses Holm correction across all 40 tests.

BaselineHigher mean for ACSignificant gain for AC
Thompson sampling8/88/8
XGBoost forecasting8/88/8
Noncyclic stock rule8/88/8
Meta-HQL, lost sales4/44/4
Meta-HQL, backlogging4/40/4
Stock rule given the true cycle0/80/8

Under backlogging, AC earned 155.4 to 164.7 more per day than Meta-HQL, with adjusted p=0.0516p = 0.0516 in each setting. The cyclic rules used the true period and targets selected in hindsight. Figure 2 shows why the period matters: under lost sales, oldest-first issuance and no fixed fee, cyclic base stock earned 474.4 per day with the true ten-day period and 331.0 to 331.4 with the five alternatives tested. AC earned 395.29 without a supplied period.

When demand is recoverable and costs are known, Table 3 gives a hybrid that retains TabPFN: the frozen model forecasts demand quantiles, and the retailer uses the cost formula to choose an order. It earned more than default AC in the tested backlogging setting with newest-first issuance and a fixed fee of 20 (Q1 below). If recurring demand is also a credible assumption, the controller in Table 6 estimates candidate periods directly from past demand. It earned 18.12 to 42.67 more per day than AC across all eight settings without being given the period or phase.

W2. How does the theory explain actual orders, and what scales?

A stock-age record helps AC distinguish orders with different immediate profits. A separate question is whether even the best order for today leaves the right stock for tomorrow. Table 14 measures both losses at inventories AC actually visited in eight small lost-sales instances. With three-day shelf life, oldest-first issuance and a fixed fee of 20, AC's order gave up 1.866 in expected immediate profit on average. Even the best one-day order gave up 3.752 against the full-horizon choice when all later orders were optimal. Thus, improving the immediate score alone would leave a measurable loss in that instance.

Lemma 6.2 connects these measurements to long-run decisions by bounding the effect of starting stock on future value under lost sales, independently of the remaining horizon. Proposition A.3 applies this bound to AC without requiring a gap between the two best immediate orders. Corollary A.4 gives conditions under which scoring, sampling and the future value forgone by one-day decisions yield sublinear regret. Proposition A.5 gives a separate sufficient test using scores, one-step rewards and inventory transitions, without computing the full optimal value function. Table 14 diagnoses a finite-horizon loss. It does not establish that the deployed scorer meets these conditions.

The deployed policy scores 601 possible orders per day over 10,000 days. Figure 4 measures about three seconds per decision at its largest tested grid of roughly 800 orders. Exact full-horizon planning was executed only on small instances. The study covers one product with zero replenishment lead time; we have not tested positive lead times or several products. For store sales, the relevant test is the perishable comparison in Table 4 (Q3 below). Figure 3 also reports lower nonperishable costs for AC than for Meta-HQL on a Rossmann trace, but the two benchmark implementations treat stock above the order target differently (Appendix G), so that comparison is not like-for-like.

Q1. How does the same frozen TabPFN perform when it forecasts demand first?

We tested three ways to turn forecasts from the same frozen TabPFN into orders. The table reports mean daily profit over 10,000 days in two settings: lost sales with oldest-first issuance (LS/FIFO) and no fixed fee, and backlogging with newest-first issuance (BL/LIFO) and a fixed fee of 20. Each FO rule is shown beside its own paired default top-ten AC result.

Forecast-to-order ruleLS/FIFO ACLS/FIFO FOBL/LIFO ACBL/LIFO FO
Median demand target (Tables 1 and 15)395.29135.95515.74528.53
Demand quantiles with known costs (Table 3)381.95374.30513.05534.59
Mean forecast with buffer 0.2 (Table 11)380.43316.04515.32532.62
Mean forecast with buffer 1.0 (Table 11)380.43424.35515.32442.47

AC outperformed median FO in all four lost-sales settings after Holm correction. Cost-optimizing FO, which evaluates orders across 129 demand quantiles, earned 21.54±4.4321.54 \pm 4.43 more per day than AC in the backlogging setting shown, where 4.43 is the half-width of the paired 95% confidence interval. Neither lost-sales comparison with cost-optimizing FO passed Holm correction. For the mean-forecast rule, raising the buffer helped under lost sales and hurt under backlogging.

Q2. Why is the sparse-prior controller competitive? Does this mean forecasting is easy?

We made this a main comparison in Table 6. The benchmark's ten cycle positions draw demand from different distributions, including a scaled beta, a lognormal and a Pareto distribution (Table 10). Both controllers receive the same recovered demands, use the same known costs and maximize the same one-day profit. Both spread an initial weight equal to one observation over the 1,801 possible capped demand values. The difference is which days they pool when estimating demand.

The comparison shows why the periodic controller was competitive. Pooling all days into one demand distribution earned 22.83 to 84.01 less per day than AC in every setting. Using the same demand observations and costs, but allowing distributions to recur at candidate periods, earned 18.12 to 42.67 more per day than AC in every setting. The controller received no period, phase or period bound, and all differences passed Holm correction within their group of eight (p<0.001p<0.001). Thus the ordering gains depended on how the controller grouped past demand. Establishing how easy demand itself was to forecast would require a direct comparison of forecast errors.

Q3. When did direct scoring win on store sales traces?

Direct scoring earned more at all ten held-out stores after both policies transferred to a new inventory regime. We simulated a three-day shelf life using 942 days of Rossmann sales per store (Table 4). On five separate stores under lost sales, FIFO and no fixed fee, we selected AC from 24 sampling and initialization configurations and FO from five safety-stock multipliers. We then fixed both configurations. FO used its selected multiplier of 1.5 and deducted all saleable stock.

Held-out store settingACSafety-stock FOAC minus FO (95% CI)Holm pp
Lost sales, FIFO, no fee1,088.891,169.20−80.31±52.78-80.31 \pm 52.780.0070.007
Backlogging, LIFO, fee of 201,159.451,037.31+122.13±34.67+122.13 \pm 34.67<0.001<0.001

The gain applies to transfer against this fixed safety-stock rule. Performance with a buffer retuned for the new regime or with cost-optimizing FO on these stores remains open.

Reviewer_KkDSOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies inventory control problem, particularly focussing on two paradigms: Action-Scoring-to-Order and Forecast-to-Order. The authors analyze the effects of each component of the two paradigms on the performance in the inventory control problem. Further theoretical analysis additionally connects the prediction error of Action-Scoring-to-Order method with online regret performance in the inventory control problem.

Critical strengths

The ablation studies are solid, and the hypotheses the studies are design to test are stated in a clear way. The theory in Section 6 is presented nicely.

Critical weaknesses

The main weakness of this work to me is that it provides limited practical guidance for practitioners to design or choose algorithms. Specifically, how can we extend the findings of the paper to a better algorithm for the inventory control problem? This paper did a solid study in identifying that various components matter in this problem, but did not, for example, provide a best performing combination of the design choices across each component.

Questions for authors

What if we use returns to go as the regression target instead of the current one-step immediate profit? And additionally, can one directly run an offline RL algorithm in this setting?

6 comments
AuthorsOpenReview ↗

Choosing an ordering policy from operating records and testing return-to-go

Thank you for asking which combination of components a retailer could put into practice. In the revised manuscript we compare complete ordering policies and test your proposed change from immediate profit to return-to-go in eight online inventory settings.

W1. Which configuration would we use in practice?

For a retailer ordering from its own operating records, we recommend full-history Decision-TabPFN, our direct action-scoring policy (AC). It uses frozen TabPFN to score each order's immediate profit from stock by age and past outcomes, then samples from the ten highest-scoring orders. If memory is limited, keeping the latest 5,000 records gave the following mean daily profits over all 10,000 days (Table 12).

Inventory settingLatest 5,000 recordsFull history
Lost sales, oldest first, no fixed fee391.23395.29
Backlogging, newest first, fixed fee of 20513.85515.74

The choice changes when the retailer can recover demand, knows its costs and has reason to expect recurring demand. Under those conditions, a controller that estimated the cycle from past demand earned 18.12 to 42.67 more per day than AC in all eight settings, without being given the period or phase (Table 6). For a retailer who wants to retain TabPFN and knows the costs, forecasting demand quantiles and optimizing one-day profit requires no safety-stock multiplier. This rule earned 21.54±4.4321.54 \pm 4.43 more per day than default AC under backlogging, newest-first issuance and a fixed fee of 20. The two lost-sales comparisons were inconclusive (Table 3).

Q1a. What happens if the model learns from return-to-go?

Immediate-profit AC earned higher mean daily profit than the online return-to-go (RTG) policy in all eight settings. The paired mean gains ranged from 62.10 to 173.57 per day, with five differences significant after Holm correction (Table 16). We matched ten 10,000-day demand paths per setting, the four initial orders, the frozen model and the top-ten sampler. Each policy used all its own completed records as its inventory evolved (Appendix L).

To form an RTG label before ordering on day tt, we summed the observed daily profits from the historical order's execution on day ii through yesterday, discounting each additional day by 0.95. Thus yesterday's order has only one observed profit, whereas an older order's label also reflects subsequent replenishments and demand shocks. These are two plausible sources of difficulty in attributing the return to the original order. For this growing-history implementation, the profit comparison supports keeping the immediate label.

Q1b. Can an offline reinforcement learning algorithm use these logs?

Yes. Once the next pre-order inventory is observed, the retailer can assemble transitions containing the inventory and backlog before ordering, the executed quantity, the realized profit and the next inventory and backlog. For instance, fitted Q-iteration could use these transitions to estimate immediate reward plus the value of later decisions. With a hidden demand cycle, a learner can also use past operating records to represent information missing from the current inventory alone.

Offline policy fitting and direct scoring use the same records differently. A fitted learner must estimate its value function or policy from a chosen batch, whereas AC reads the examples at decision time with its pretrained weights fixed. AC can begin on day 5 with three completed examples and use 9,998 by day 10,000. Table 16 compares two labels within this frozen-model protocol. It supports immediate-profit scoring for that use, while the available transitions make a separate offline RL comparison feasible.

Thanks for the detailed response. However, my main question remains: it still occurs to me that the paper does not provide a principled way to decide which algorithm to use given a scenario - only experimental results in one setting are presented. Therefore, I would like to ask for general principles/intuitions of when one method (e.g. AC) can outperform the others (e.g. FO, offline RL) and the reasons.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.