Pith. sign in

REVIEW 4 major objections 4 minor 30 references

All twelve continuous-time generative models are special cases of one mean-field game, and the paper turns that equivalence into an executable API in which a four-part cost tuple is the only user-facing specification.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 03:48 UTC pith:6VCD6DDS

load-bearing objection The cost-tuple framework is genuinely useful, but the paper's own numbers fail to support the 'lossless' claim for interaction terms, and DI-Flow's headline result is confounded — worth engaging, not accepting as-is. the 4 major comments →

arxiv 2607.23026 v1 pith:6VCD6DDS submitted 2026-07-25 cs.LG cs.AI

All in One: Generative Modeling as Mean-Field Game Design

Kun Zhao , Xu Chen This is my paper
classification cs.LG cs.AI
keywords mean-field gamesgenerative modelingcost tuplenormalizing flowsSchrödinger bridgescore-based modelsinteraction termmode coverage
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper tries to establish that continuous-time generative modeling is, at bottom, a mean-field game-design problem: pick a terminal cost, a running cost, an interaction term, and a noise level, and the transport of data samples to a reference distribution is the equilibrium of a forward-backward system of PDEs. Twelve existing models, from continuous normalizing flows to Schrödinger bridges, are claimed to be exactly the special cases yielded by different choices of these four components. The paper's library makes the cost tuple the only user-facing specification, so all twelve models share one training loop, one log-Jacobian computation, and one sampler. It also proposes DI-Flow, a new interaction term based on a differentiable entropy estimate that pushes particles apart and improves coverage on a multimodal benchmark, and shows that learning-based mean-field-game solvers outperform gradient-based neural training on stochastic-dynamics rows. A sympathetic reader would care because this recasts model design as choosing a cost, and model comparison as comparing costs, rather than reimplementing each algorithm.

Core claim

The paper's central discovery is an executable equivalence: generative modeling can be posed as the variational mean-field-game problem of minimizing a terminal cost plus integrated running and interaction costs, subject to a Fokker-Planck constraint. In this formulation, the optimal velocity field and population density are the joint solution of a forward Fokker-Planck equation and a backward Hamilton-Jacobi-Bellman equation coupled through the variational derivative of the interaction term. The paper claims every model in its twelve-model taxonomy corresponds to one tuple (M, I, L, σ), and that when the cost tuple is implemented as composable plug-ins, the same training loop, log-Jacobian,

What carries the argument

The central object is the cost tuple (M, I, L, σ)—terminal cost, interaction functional, running cost, noise level—inserted into the mean-field variational objective. The argument runs through the coupled FPK-HJB optimality system: the forward equation propagates the particle density under the current velocity, the backward equation propagates the value function, and the variational derivative δI/δρ couples them. The paper's training algorithm evaluates the interaction directly on the mini-batch as a particle loss with an accumulated log-Jacobian, and the DI-Flow design instantiates I as the logarithm of a kernel density estimate, whose gradient acts as a repulsive force. Value iteration and

Load-bearing premise

The particle objective used in training evaluates the interaction term on a mini-batch, while the mean-field equilibrium is defined through the variational derivative of a population functional; the paper does not establish that optimizing the batch objective converges to the same equilibrium as the FPK-HJB system.

What would settle it

Train DI-Flow and OT-Flow with identical network capacity on a six-mode Gaussian mixture across a range of batch sizes, and compare against the grid-based mean-field solver on the same cost tuple. If DI-Flow's coverage advantage vanishes at matched capacity, or if its behavior fails to approach the equilibrium solver as batch size grows, the claim that the KDE interaction implements a genuine mean-field design is not supported.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Any differentiable population objective can be plugged in as the interaction term I without altering the training loop; the paper describes this design space as vast and largely unexplored.
  • The learning-based mean-field-game solvers solve stochastic-dynamics rows (Schrödinger bridge, generalized Schrödinger bridge, stochastic OT-normalizing flow) with near-perfect coverage in under a second, while neural training on the same rows collapses.
  • DI-Flow, which differs from OT-Flow only by its nonzero interaction term, achieves the best KDE log-likelihood among deterministic models on the Ring GMM benchmark.
  • Because all rows share the same training loop and sampler, differences in model behavior can be traced directly to the four cost functions, making model comparison a comparison of costs.
  • The same variational structure is expected to accommodate richer interaction geometries, such as graphon-style heterogeneous networks, since the coupling enters only through the interaction functional.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Editorial inference: the paper's DI-Flow comparison is confounded by a larger network, and the paper itself flags in Section 4.1 that a matched-capacity ablation is needed; until that ablation, the reported coverage gain cannot be attributed solely to the KDE interaction.
  • Editorial inference: if the mini-batch interaction does not track the population functional as batch size grows, DI-Flow's mode-covering behavior may be a finite-batch artifact rather than a genuine mean-field equilibrium; a batch-size scaling test would separate the two.
  • Editorial inference: the inverse-design discussion implies that observed generative trajectories could be reverse-engineered into the cost tuple that rationalizes them, turning model selection into an inverse mean-field-game problem; the paper does not implement this.
  • Editorial inference: the cost-tuple abstraction suggests an automated cost-search loop—propose an interaction, train, evaluate, and refine—which the paper gestures at but does not run; this is a testable next step.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes MFGLab, a PyTorch library that unifies twelve continuous-time generative models under a mean-field-game cost tuple (M, I, L, sigma). The authors claim that this cost tuple is the only user-facing specification, with a shared training loop, log-Jacobian, and reverse-ODE sampler. They introduce DI-Flow, a new interaction term based on a differentiable KDE entropy functional, and report that it achieves the best KDE log-likelihood among deterministic baselines on a Ring GMM benchmark. They also contribute learning-based MFG solvers (VI and MFG-AC) and claim these substantially outperform neural training on stochastic-dynamics rows. The experiments compare the unified API against hand-coded implementations on two 2D benchmarks.

Significance. If the cost-tuple abstraction were truly lossless and DI-Flow's interaction term were shown to implement the mean-field equilibrium, this would be a valuable contribution: it would turn a heterogeneous collection of generative models into a single executable design space with a principled way to add population-level objectives. The library idea is attractive and the connection to MFG theory is timely. However, the current evidence does not establish the central claims: the DI-Flow Lab/Individual discrepancy directly contradicts the lossless claim, the batch-evaluated interaction is not linked to the variational derivative in the MFG equilibrium, the headline DI-Flow result is confounded by a larger network, and the reported experiments are single runs without error bars. These issues are load-bearing and require substantial revision.

major comments (4)
  1. [§4.1, Table 2 (DI-Flow row)] The central claim that 'Lab matches Individual across all rows' is contradicted by the DI-Flow row: Lab gives Cov 0.997 and KDE -3.42, while Individual gives Cov 0.484 and KDE -7.84, and the MFG Solver gives Cov 0.430 and KDE -5.17. The authors dismiss this as 'different random initialisation ordering', but no evidence supports that. This discrepancy means the lossless-API claim fails for the one row where the new method is introduced, and the 'best KDE among sigma=0 baselines' claim is therefore not established by the reported numbers.
  2. [§2.2 Eq. (4) and Algorithm 2, line 8] The MFG equilibrium in Eq. (3) is defined through the variational derivative delta I / delta rho of a population functional, but Algorithm 2 evaluates I(x_t, {x_j}) directly in the per-particle loss. The paper states that 'any differentiable functional of the mini-batch can be plugged in as I' without deriving or testing whether the gradient of the batch-evaluated objective equals delta I / delta rho. No argument or experiment shows that optimizing this finite-batch objective converges to the FPK-HJB equilibrium. DI-Flow's mode-covering behavior may therefore be an artifact of the batch estimator rather than a genuine mean-field game design.
  3. [§3.3 and §4] The abstract and contribution list claim 'learning-based MFG solvers' that 'substantially outperform neural training', but the experiments in Tables 2 and 3 only compare the unified API, hand-coded classes, and VI (grid-based backward DP). MFG-AC is described in Appendix B but no experimental results are reported for it. Thus the 'substantially outperform' claim is only supported for a non-learning grid solver on 2D problems, not for the learning-based solver that is a stated contribution.
  4. [§4 and Appendix C] All quantitative results are single runs with no seeds, no error bars, and no statistical significance tests. Given the small margins in some comparisons (e.g., OT-Flow Lab vs Individual on Ring GMM) and the large DI-Flow discrepancy, this is insufficient to support the paper's strong conclusions. The authors should report multiple seeds and ideally confidence intervals, especially for the headline DI-Flow result.
minor comments (4)
  1. [Algorithm 1, line 9] The algorithm uses delta I / delta rho in the backward step, but the paper never defines how this derivative is computed or estimated in practice. This is particularly important because the training algorithm in Algorithm 2 uses I directly, not its variational derivative.
  2. [§3.2, Eq. (6)] The KDE bandwidth h depends on sigma_hat_t, but sigma_hat_t is not defined. It should be specified whether this is the empirical standard deviation of the current mini-batch or of the population estimate.
  3. [Table 2 footnote] The DI-Flow Lab training time of 601 s vs OT-Flow 41 s is attributed to O(N^2) KDE interaction; this should be stated in the main text, not only in Appendix A, since it matters for the practical claim.
  4. [Figure 3] Figure 3 is not explicitly described in the text; it would help to annotate which panel corresponds to DI-Flow and how the visualizations relate to the quantitative metrics.

Circularity Check

0 steps flagged

No significant circularity: the taxonomy is externally cited, DI-Flow's interaction is a stated cost choice rather than a fitted target, and no equation defines its outputs in terms of its inputs.

full rationale

The paper's load-bearing unification claim is explicitly imported from the external taxonomy of ref. [27] (Zhang and Katsoulakis, different authors), not from the present authors' own prior work. Changing only the cost tuple to recover existing models is a citation-based mapping, not a derivation that re-imports the paper's own conclusions. The empirical claim that the MFGLab API is lossless is tested by comparing the Lab implementation against dedicated Individual implementations on fixed cost tuples; this is a measurement, not a definition. DI-Flow's interaction term I = γ log ρ̂_t is a proposed cost design with a freely chosen strength γ, and the paper explicitly admits that its superior KDE score is confounded by a larger network, so the result is not presented as a forced consequence of a fitted parameter. Algorithm 2's use of the batch-evaluated I directly in the particle loss, rather than the variational derivative δI/δρ from Eq. (3), is a correctness/validation gap (the training objective may not solve the stated MFG equilibrium), but it is not circular: the reported numbers simply measure that explicitly defined training objective. The self-citations by co-author Xu Chen ([4], [18], [28], [29]) appear only as related solver literature and inverse-design context; they are not invoked to justify the central cost-tuple taxonomy or to forbid alternatives. No uniqueness theorem imported from the authors, no fitted parameter renamed as prediction, and no equation reduces to its own input by construction. The paper's own admitted limitations—DI-Flow Lab vs Individual coverage mismatch and missing matched-capacity ablation—are empirical weaknesses, not circularity.

Axiom & Free-Parameter Ledger

1 free parameters · 5 axioms · 0 invented entities

The central claims rest on the prior MFG taxonomy for the twelve-model mapping, on a standard normalizing-flow particle objective for I=0, and on an unproved batch-evaluation approximation for the new interaction term. One free parameter, γ, controls the DI-Flow interaction strength and is not reported. No new physical or mathematical entities are postulated.

free parameters (1)
  • γ (DI-Flow interaction strength) = not reported
    Introduced in Eq. (6); controls the magnitude of the repulsive KDE entropy penalty. It is not listed in the hyperparameter table and no sensitivity analysis is given, so its value is effectively chosen to make DI-Flow work.
axioms (5)
  • domain assumption The potential-MFG variational problem (Eq. 1) with FPK constraint (2) and HJB optimality (3) is the correct unifying formulation for the listed generative models.
    The paper takes this MFG formalization as given, citing standard MFG theory and ref [27]; it is not proved here.
  • domain assumption Each row of Table 1 correctly maps a named generative model to a cost tuple, as reproduced from ref [27].
    The paper relies entirely on the prior taxonomy for the 'twelve models in one variational problem' claim and does not re-derive or numerically validate the mapping for all rows.
  • standard math For σ=0 and I=0, the per-particle loss (4) with accumulated log-Jacobian equals the population objective.
    This is the standard normalizing-flow change-of-variables argument used in CNFs and OT-Flow; it is correct under mild regularity conditions.
  • ad hoc to paper Batch evaluation of I(x_t, {x_j}) in Algorithm 2 gives the same dynamics as the population functional's variational derivative δI/δρ in the HJB equation (3).
    The paper introduces the interaction term as a differentiable batch functional but does not prove or test that the particle objective has the same equilibrium as the mean-field game. This is the load-bearing approximation behind DI-Flow.
  • domain assumption The mini-batch KDE with Silverman bandwidth is a faithful estimator of ρ_t for the purpose of the interaction term.
    DI-Flow's repulsion is defined through a finite-sample KDE; the paper does not analyze the bias or the effect of batch size on the equilibrium.

pith-pipeline@v1.3.0-alltime-deepseek · 12010 in / 11159 out tokens · 114579 ms · 2026-08-01T03:48:58.792824+00:00 · methodology

0 comments
read the original abstract

Mean-field games (MFGs) offer a unifying lens on continuous-time generative modeling: a cost tuple recovering twelve prominent models---Continuous Normalizing Flows, OT-Flow, Score-based Models, Schr\"{o}dinger Bridges, and more---as special cases of one variational problem. Yet two dimensions of this space remain entirely unexplored: the interaction term $\mathcal{I}$ is set to zero in many existing models, and the rich family of MFG solvers has never been applied to generative modeling. We address both gaps with MFGLab an open-source PyTorch library whose primary API is the cost tuple: all twelve models are specified by four composable cost functions, and the training loop, log-Jacobian, and reverse-ODE sampler are shared automatically. We additionally propose DI-Flow, a novel cost design that uses a differentiable entropy functional to encourage mode coverage, and provide learning-based MFG solvers that substantially outperform neural training on stochastic-dynamics rows. Experiments on two 2-D benchmarks confirm that the unified API is lossless relative to hand-coded implementations.

Figures

Figures reproduced from arXiv: 2607.23026 by Kun Zhao, Xu Chen.

Figure 1
Figure 1. Figure 1: Overview of this paper’s three contributions. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: MFGLAB framework overview. The cost tuple (M, I,L, σ) fully determines the training objective and reverse-ODE sampler; the same particle simulation, Adam loop, and generation procedure are shared across all twelve models in [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Generated samples on Ring GMM (K=6, 3 000 iterations, batch 512) [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Value field V (x, t) and optimal velocity ∥u ∗ (x, 0)∥ computed by VI (OT-Flow row, λ=1, σ=0), illustrating how the HJB solution propagates backward from terminal cost to initial velocity. Left: terminal cost V (x, T)=− log π(x). Centre: value-to-go V (x, 0), propagated backward via HJB. Right: initial speed field with velocity arrows. 6 Conclusion We presented MFGLAB, a unified and accessible framework th… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

30 extracted references · 3 linked inside Pith

  1. [1]

    Pieter Abbeel and Andrew Y . Ng. Apprenticeship learning via inverse reinforcement learning. InInternational Conference on Machine Learning, pages 1–8, 2004

    2004

  2. [2]

    Albergo, Nicholas M

    Michael S. Albergo, Nicholas M. Boffi, and Eric Vanden-Eijnden. Stochastic interpolants: A unifying framework for flows and diffusions.Journal of Machine Learning Research, 26, 2025

    2025

  3. [3]

    Ricky T. Q. Chen, Yulia Rubanova, Jesse Bettencourt, and David Duvenaud. Neural ordinary differential equations.Advances in Neural Information Processing Systems, 31, 2018

    2018

  4. [4]

    Learning dual mean field games on graphs

    Xu Chen, Shuo Liu, and Xuan Di. Learning dual mean field games on graphs. InEuropean Conference on Artificial Intelligence, 2023

    2023

  5. [5]

    SVGD as a kernelized wasserstein gradient flow of the chi-squared divergence.Advances in Neural Information Processing Systems, 2020

    Sinho Chewi, Thibaut Le Gouic, Chen Lu, Tyler Maunu, and Philippe Rigollet. SVGD as a kernelized wasserstein gradient flow of the chi-squared divergence.Advances in Neural Information Processing Systems, 2020

    2020

  6. [6]

    nflows: normalizing flows in PyTorch, 2020

    Conor Durkan, Artur Bekasov, Iain Murray, and George Papamakarios. nflows: normalizing flows in PyTorch, 2020

    2020

  7. [7]

    Will Grathwohl, Ricky T. Q. Chen, Jesse Bettencourt, Ilya Sutskever, and David Duvenaud. FFJORD: Free-form continuous dynamics for scalable reversible generative models.Interna- tional Conference on Learning Representations, 2019

    2019

  8. [8]

    Learning mean-field games

    Xin Guo, Anran Hu, Renyuan Xu, and Junzi Zhang. Learning mean-field games. InAdvances in Neural Information Processing Systems, volume 32, 2019

    2019

  9. [9]

    Generative adversarial imitation learning

    Jonathan Ho and Stefano Ermon. Generative adversarial imitation learning. InAdvances in Neural Information Processing Systems, volume 29, 2016

    2016

  10. [10]

    Denoising diffusion probabilistic models.Advances in Neural Information Processing Systems, 33:6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models.Advances in Neural Information Processing Systems, 33:6840–6851, 2020

    2020

  11. [11]

    Unsupervised solution operator learning for mean-field games via sampling-invariant parametrizations.arXiv preprint arXiv:2401.15482, 2024

    Han Huang and Rongjie Lai. Unsupervised solution operator learning for mean-field games via sampling-invariant parametrizations.arXiv preprint arXiv:2401.15482, 2024

    Pith/arXiv arXiv 2024

  12. [12]

    Malhamé, and Peter E

    Minyi Huang, Roland P. Malhamé, and Peter E. Caines. Large population stochastic dynamic games: Closed-loop McKean-Vlasov systems and the Nash certainty equivalence principle. Communications in Information and Systems, 6(3):221–252, 2006

    2006

  13. [13]

    Mean field games.Japanese Journal of Mathematics, 2(1):229–260, 2007

    Jean-Michel Lasry and Pierre-Louis Lions. Mean field games.Japanese Journal of Mathematics, 2(1):229–260, 2007

    2007

  14. [14]

    Learning mean field games: A survey

    Mathieu Laurière, Sarah Perrin, Matthieu Geist, and Olivier Pietquin. Learning mean field games: A survey. InarXiv preprint arXiv:2205.12944, 2022

    Pith/arXiv arXiv 2022

  15. [15]

    Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling.International Conference on Learning Representations, 2023

    2023

  16. [16]

    Let us build bridges: Understanding and extending diffusion generative models

    Xingchao Liu, Lemeng Wu, Mao Ye, and Qiang Liu. Let us build bridges: Understanding and extending diffusion generative models. InNeurIPS Workshop on Score-Based Methods, 2022

    2022

  17. [17]

    Sliced-wasserstein flows: Nonparametric generative modeling via optimal transport and diffu- sions.International Conference on Machine Learning, 2019

    Antoine Liutkus, Umut Simsekli, Szymon Majewski, Alain Durmus, and Fabian-Robert Stöter. Sliced-wasserstein flows: Nonparametric generative modeling via optimal transport and diffu- sions.International Conference on Machine Learning, 2019

    2019

  18. [18]

    Zhaobin Mo, Xu Chen, Xuan Di, Elisa Iacomini, Chiara Segala, Michael Herty, and Mathieu Lauriere. A game-theoretic framework for generic second-order traffic flow models using mean field games and adversarial inverse reinforcement learning.Transportation Science, 58(6):1403–1426, 2024

    2024

  19. [19]

    Boltzmann generators: Sampling equilibrium states of many-body systems with deep learning.Science, 365(6457), 2019

    Frank Noé, Simon Olsson, Jonas Köhler, and Hao Wu. Boltzmann generators: Sampling equilibrium states of many-body systems with deep learning.Science, 365(6457), 2019. 10

    2019

  20. [20]

    OT-Flow: Fast and accurate continuous normalizing flows via optimal transport.AAAI Conference on Artificial Intelligence, 2021

    Derek Onken, Samy Wu Fung, Xingjian Li, and Lars Ruthotto. OT-Flow: Fast and accurate continuous normalizing flows via optimal transport.AAAI Conference on Artificial Intelligence, 2021

    2021

  21. [21]

    Fictitious play for mean field games: Continuous time analysis and applications

    Sarah Perrin, Julien Pérolat, Mathieu Laurent, Matthieu Geist, Romuald Elie, and Olivier Pietquin. Fictitious play for mean field games: Continuous time analysis and applications. In Advances in Neural Information Processing Systems, volume 33, 2020

    2020

  22. [22]

    On imitation in mean-field games

    Giorgia Ramponi, Pavel Kolev, Olivier Pietquin, Niao He, Mathieu Lauriere, and Matthieu Geist. On imitation in mean-field games. InAdvances in Neural Information Processing Systems, volume 36, 2023

    2023

  23. [23]

    Osher, Wuchen Li, Levon Nurbekyan, and Samy Wu Fung

    Lars Ruthotto, Stanley J. Osher, Wuchen Li, Levon Nurbekyan, and Samy Wu Fung. A machine learning framework for solving high-dimensional mean field game and mean field control problems.Proceedings of the National Academy of Sciences, 117(17):9183–9193, 2020

    2020

  24. [24]

    Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole

    Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. International Conference on Learning Representations, 2021

    2021

  25. [25]

    normflows: A PyTorch package for nor- malizing flows, 2023

    Vincent Stimper, David Liu, Andrew Campbell, Vincent Berenz, Lukas Ryll, Bernhard Schölkopf, and José Miguel Hernández-Lobato. normflows: A PyTorch package for nor- malizing flows, 2023

    2023

  26. [26]

    Diffusers: State-of-the-art diffusion models, 2022

    Patrick von Platen, Suraj Patil, Anton Lozhkov, Pedro Cuenca, Nathan Lambert, Kashif Rasul, Tim Dettmers, and Thomas Wolf. Diffusers: State-of-the-art diffusion models, 2022

    2022

  27. [27]

    Zhang and Markos A

    Benjamin J. Zhang and Markos A. Katsoulakis. A mean-field games laboratory for generative modeling.arXiv preprint arXiv:2304.13534, 2023

    Pith/arXiv arXiv 2023

  28. [28]

    Stochastic semi-gradient descent for learning mean field games with population-aware function approximation

    Chenyu Zhang, Xu Chen, and Xuan Di. Stochastic semi-gradient descent for learning mean field games with population-aware function approximation. InInternational Conference on Learning Representations, 2025

    2025

  29. [29]

    Graphon mean field games with a representative player: Analysis and learning algorithm

    Fuzhong Zhou, Chenyu Zhang, Xu Chen, and Xuan Di. Graphon mean field games with a representative player: Analysis and learning algorithm. InInternational Conference on Machine Learning, 2024

    2024

  30. [30]

    Ziebart, Andrew Maas, J

    Brian D. Ziebart, Andrew Maas, J. Andrew Bagnell, and Anind K. Dey. Maximum entropy inverse reinforcement learning. InAAAI Conference on Artificial Intelligence, pages 1433–1438, 2008. 11 A Experimental Setup A 2-D six-component Ring Gaussian Mixture ( K=6, radius 2.5, std 0.4). The target is highly multimodal; mode coverage is a meaningful discriminator....

    2008