papers/prioritized-generative-replay/sau.json

{
  "paper_id": "prioritized-generative-replay",
  "paper_title": "Prioritized Generative Replay",
  "D1": [
    {
      "id": "prioritized-generative-replay-D1-001",
      "claim": "Standard online RL training: real and synthetic replay buffers each hold up to 1,000,000 transitions; 100K environment step interaction budget; baseline UTD ratio 20; baseline batch size 256 with synthetic data ratio r=0.5 (128 real + 128 generated transitions per batch).",
      "source": "Section 4.3 paragraph 6; Section 5 paragraphs 1-2; Section 5.3 paragraph 2"
    },
    {
      "id": "prioritized-generative-replay-D1-002",
      "claim": "Probability of randomly dropping the relevance function condition y during diffusion model training to enable classifier-free guidance at sampling: 0.25.",
      "source": "Section 5, paragraph 2"
    },
    {
      "id": "prioritized-generative-replay-D1-003",
      "claim": "Fraction of policy gradient steps at which the curiosity head (ICM forward dynamics relevance function) is updated, with remaining steps updating the policy only: 0.05.",
      "source": "Section 5, paragraph 4"
    },
    {
      "id": "prioritized-generative-replay-D1-004",
      "claim": "Frequency at which the inner loop (diffusion model retraining and synthetic data regeneration) executes; determined via elbow method on hopper-stand: once every 10,000 environment steps.",
      "source": "Appendix D, paragraph 5"
    },
    {
      "id": "prioritized-generative-replay-D1-005",
      "claim": "DMC tasks (state-based and pixel-based): standard 100K environment step budget for all tasks except finger-turn-hard, which uses an extended 300K budget due to sparser rewards; 5 random seeds for both state-based and pixel-based DMC experiments.",
      "source": "Section 5, Table 1 caption"
    },
    {
      "id": "prioritized-generative-replay-D1-006",
      "claim": "OpenAI Gym tasks (Walker2d-v2, HalfCheetah-v2, Hopper-v2): 100K environment step budget; 3 random seeds.",
      "source": "Section 5, Table 2 caption"
    },
    {
      "id": "prioritized-generative-replay-D1-007",
      "claim": "DMLab randomized 3D maze environment: 84x84 first-person RGB observations; 9 discrete actions at repeat frequency 4; max 1800 steps per episode; sparse reward +10 at goal (0 otherwise); total 10,000,000 environment steps; 10 random seeds.",
      "source": "Appendix A.2, paragraph 1; Appendix A.2, Table 5 caption"
    },
    {
      "id": "prioritized-generative-replay-D1-008",
      "claim": "DMLab noisy-TV injection: the lower-right 42x42 pixel quadrant of each 84x84 first-person observation is replaced with per-pixel noise sampled uniformly from [0, 255], independently across pixels and timesteps.",
      "source": "Appendix A.2, paragraph 2"
    },
    {
      "id": "prioritized-generative-replay-D1-009",
      "claim": "Noisy dynamics experiments on DMC: 100K environment step budget; at each training step, 20% of transitions per batch receive isotropic Gaussian noise on states; 5 random seeds.",
      "source": "Appendix C, paragraph 2; Appendix C, Table 8 caption"
    },
    {
      "id": "prioritized-generative-replay-D1-010",
      "claim": "Baseline policy MLP architecture used in all standard experiments (prior to scaling): 2 hidden layers of width 256.",
      "source": "Section 5.3, paragraph 1"
    },
    {
      "id": "prioritized-generative-replay-D1-011",
      "claim": "Scaled policy MLP architecture for scaling experiments: 3 hidden layers of width 512, resulting in approximately 6x more parameters than the baseline 2x256 configuration.",
      "source": "Section 5.3, paragraph 1"
    },
    {
      "id": "prioritized-generative-replay-D1-012",
      "claim": "Model parameter counts (Table 3): PGR generative model (diffusion plus curiosity head) 7,390,000 parameters (+3.7% over SYNTHER); SYNTHER unconditional generative model 7,120,000; REDQ baseline policy 9,860,000.",
      "source": "Section 5, Table 3"
    },
    {
      "id": "prioritized-generative-replay-D1-013",
      "claim": "GPU VRAM for synthetic data generation (Table 3): PGR requires 6.67 GB (54.7% more than SYNTHER at 4.31 GB); both fit easily on modern GPUs with an upper bound of 12 GB.",
      "source": "Section 5, Table 3; Section 5, Table 3 caption"
    },
    {
      "id": "prioritized-generative-replay-D1-014",
      "claim": "RND relevance function architecture: 3-layer CNNs with bottleneck latent dimension 64 and feature output dimension 512, followed by a 2-layer MLP projection also of dimension 512.",
      "source": "Appendix A.1, paragraph 2"
    },
    {
      "id": "prioritized-generative-replay-D1-015",
      "claim": "CTS density model: visual observations resized to 42x42 pixels with 8 context bins; relevance function F = (N_hat(s,a) + 0.01)^(-1/2), where pseudo-count epsilon 0.01 avoids division by zero.",
      "source": "Appendix A.1, paragraph 3; Appendix A.1, Eq (7)"
    },
    {
      "id": "prioritized-generative-replay-D1-016",
      "claim": "ECO relevance function: ResNet-18 embedder output dim 512, 4-layer MLP dim 512; F(s) = 0.03 * (0.5 - percentile-90(C(E(s), E(s_i)))) for all s_i in memory buffer M of size 200; comparator C trained via logistic regression to estimate reachability between observation embeddings.",
      "source": "Appendix A.2, paragraph 5"
    },
    {
      "id": "prioritized-generative-replay-D1-017",
      "claim": "Scaling experiments: scaled UTD ratio 40 (doubled from 20); synthetic buffer capacity doubled to 2,000,000; synthetic data ratio variants r=0.75 with batch 512 and r=0.875 with batch 1024; real transitions fixed at 128 per batch across all ratios; 3 random seeds.",
      "source": "Section 5.3, paragraphs 1-4; Section 5.3, Figure 7 caption"
    },
    {
      "id": "prioritized-generative-replay-D1-018",
      "claim": "MSE dynamics faithfulness analysis: performed at epoch 50 (halfway through online policy learning); measures mean-squared error between generated and ground-truth next states and rewards over 10,000 generated transitions across 3 OpenAI Gym environments.",
      "source": "Section 5.2, paragraph 3"
    },
    {
      "id": "prioritized-generative-replay-D1-019",
      "claim": "Curiosity F-value and tSNE analyses: distribution of curiosity F-values (Eq. 5) measured over 10,000 real transitions every 10,000 timesteps; tSNE projection of 10,000 generated transitions for visualization in Figure 2.",
      "source": "Section 5.2, paragraph 6; Section 4, Figure 2 caption"
    },
    {
      "id": "prioritized-generative-replay-D1-020",
      "claim": "Exploration bonus experiments: intrinsic curiosity reward weight 0.1 when added as exploration bonus to REDQ or SYNTHER extrinsic reward; 3 random seeds for both explicit (Table 6) and implicit (Table 7) exploration bonus experiments on DMC-100K.",
      "source": "Appendix B.1, paragraph 1; Appendix B.2, Table 6 caption"
    },
    {
      "id": "prioritized-generative-replay-D1-021",
      "claim": "Pixel-based observations for DMC and DMLab tasks: 3 RGB channels; for pixel-based DMC, synthetic data is generated in the latent space of the policy CNN visual encoder rather than in raw pixel space.",
      "source": "Section 5, paragraph 3"
    },
    {
      "id": "prioritized-generative-replay-D1-022",
      "claim": "VAE ablation architecture (replacing diffusion model): residual net encoder with 4 layers and decoder with 8 layers; each layer has bottleneck dimension 128; latent space dimension 32; total parameters designed to match SYNTHER/PGR diffusion model capacity of approximately 6.8M.",
      "source": "Appendix D, paragraph 3"
    },
    {
      "id": "prioritized-generative-replay-D1-023",
      "claim": "VAE ablation parameter counts: encoder 2,600,000 parameters; decoder 5,300,000 parameters; approximate total 6,800,000 parameters, designed to match SYNTHER/PGR diffusion model capacity for fair comparison.",
      "source": "Appendix D, paragraph 3"
    }
  ],
  "D2": [
    {
      "id": "prioritized-generative-replay-D2-001",
      "claim": "Classifier-Free Guidance Diffusion Training Loss: L_diff(θ) = E_t,x_0,ε [||ε - ε_θ(x_t, t, c)||² + w·||ε - ε_θ(x_t, t, ∅)||²] for conditional and unconditional joint training",
      "source": "Section 3, Eq (2)"
    },
    {
      "id": "prioritized-generative-replay-D2-002",
      "claim": "Classifier-Free Guidance Sampling: x_{t-1} = (1+w)·ε_θ(x_t, t, c) - w·ε_θ(x_t, t, ∅) with guidance weight w controlling condition strength",
      "source": "Section 3, paragraph 4"
    },
    {
      "id": "prioritized-generative-replay-D2-003",
      "claim": "Return-Based Relevance Function: R_return(τ) = Σ_{t} γ^t r_t, states with higher discounted return sum are more relevant for replay",
      "source": "Section 4.2, Eq (3)"
    },
    {
      "id": "prioritized-generative-replay-D2-004",
      "claim": "TD-Error Relevance Function: R_TD(s,a) = |r + γ·max_{a'}Q(s',a') - Q(s,a)|, higher TD error → higher relevance for generation",
      "source": "Section 4.2, Eq (4)"
    },
    {
      "id": "prioritized-generative-replay-D2-005",
      "claim": "Curiosity-Based Relevance Function (Eq. 5): F(s,a,s',r) = 1/2||g(h(s),a) - h(s')||^2, where h is a learned feature encoder mapping states to a latent space, g is a forward dynamics model predicting next-state features from current latent features and action; higher prediction error signals greater novelty and thus higher priority for conditional generation.",
      "source": "Section 4.2, Eq (5)"
    },
    {
      "id": "prioritized-generative-replay-D2-006",
      "claim": "ICM Loss (Eq. 5): L_ICM = 1/2||g(h(s_t),a_t) - h(s_{t+1})||^2 minimizes latent forward-dynamics prediction error over real transitions; encoder h and forward model g are jointly optimized so the prediction error serves as both training objective and curiosity relevance score F(s,a,s',r). Updated on only 5% of policy gradient steps (Section 5 paragraph 4).",
      "source": "Section 4.2, Eq (5) context; Section 5 paragraph 4"
    },
    {
      "id": "prioritized-generative-replay-D2-007",
      "claim": "RND Relevance Function (Eq. 6): F(s,a,s',r) = 1/2||f_theta(s') - f(s')||^2, where f is a fixed randomly-initialized neural network (CNN+MLP, dim 512) and f_theta is a trainable predictor network of identical architecture; the MSE between their next-state embeddings measures novelty — higher error indicates less-frequently visited states and thus higher relevance for generation.",
      "source": "Appendix A.1, Eq (6)"
    },
    {
      "id": "prioritized-generative-replay-D2-008",
      "claim": "RND Predictor Training Loss: L_RND = ||f̂(s_t) - f(s_t)||² where f is a fixed random network and f̂ is trained to predict f's output; prediction error = exploration bonus",
      "source": "Appendix A.1, Eq (6) context"
    },
    {
      "id": "prioritized-generative-replay-D2-009",
      "claim": "CTS Pseudo-Count Relevance Function: N̂(s) ≈ (e^{||f̂(s)-f(s)||²/2σ²})/(1 - e^{||f̂(s)-f(s)||²/2σ²}), pseudo-counts derived from context tree switching density model",
      "source": "Appendix A.1, Eq (7)"
    },
    {
      "id": "prioritized-generative-replay-D2-010",
      "claim": "Episodic Curiosity (ECO) Relevance Function: R_ECO(s_t) = ||h(s_t) - h_EMA(s_t)||² where h is a trainable embedding and h_EMA its exponential moving average, rewarding novelty within episode",
      "source": "Appendix A.2, Eq (8)"
    },
    {
      "id": "prioritized-generative-replay-D2-011",
      "claim": "Noisy Networks Layer Transformation: y = (μ_w + σ_w⊙ε_w)·x + (μ_b + σ_b⊙ε_b) with learnable μ,σ and fixed noise ε~N(0,I) for exploration via parameter perturbation",
      "source": "Appendix B.2, Eq (9)"
    },
    {
      "id": "prioritized-generative-replay-D2-012",
      "claim": "PGR Outer+Inner Loop: Outer loop samples replay state s~B, inner loop generates trajectory from s via diffusion model, evaluates relevance R(τ), trains generator on high-R trajectories",
      "source": "Section 4.2, Algorithm 1; Section 4.3, paragraph 4"
    },
    {
      "id": "prioritized-generative-replay-D2-013",
      "claim": "Bootstrapped Q-Value Exploration: Q_k(s,a) = base Q-network with k-th bootstrap mask, Thompson sampling selects argmax_a Q_k(s,a) for deep exploration",
      "source": "Appendix B.2, paragraph 2"
    },
    {
      "id": "prioritized-generative-replay-D2-014",
      "claim": "Conditional Prompting Strategy for Generation: c = concat(s_t, a_{t-1}, h_t) where s_t is current state, a_{t-1} previous action, h_t optional history embedding; diffusion model generates s_{t+1}|c",
      "source": "Section 4.3, paragraph 4"
    }
  ],
  "D3": [
    {
      "id": "prioritized-generative-replay-D3-001",
      "claim": "Evaluate PGR variants (TD-Error, Reward, Curiosity) against model-free baselines (SAC, REDQ), model-based baselines (MBPO, DREAMER-V3), and unconditional generative replay (SYNTHER) on four state-based DeepMind Control Suite tasks (quadruped-walk, cheetah-run, reacher-hard, finger-turn-hard) over 100K environment steps (300K for finger-turn-hard), measuring average return over 5 random seeds.",
      "source": "Section 5.1, Table 1 (state)"
    },
    {
      "id": "prioritized-generative-replay-D3-002",
      "claim": "Evaluate PGR variants (TD-Error, Reward, Curiosity) against DRQ-V2 and unconditional generative replay (SYNTHER) on two pixel-based DeepMind Control Suite tasks (walker-walk, cheetah-run) over 100K environment steps, with synthetic data generated in the latent space of the policy's CNN visual encoder, measuring average return over 5 random seeds.",
      "source": "Section 5.1, Table 1 (pixel)"
    },
    {
      "id": "prioritized-generative-replay-D3-003",
      "claim": "Evaluate PGR variants (TD-Error, Reward, Curiosity) against SAC, REDQ, MBPO, DREAMER-V3, and SYNTHER on three OpenAI Gym state-based continuous control tasks (Walker2d-v2, HalfCheetah-v2, Hopper-v2) over 100K environment steps, measuring average return over 3 random seeds.",
      "source": "Section 5.1, Table 2"
    },
    {
      "id": "prioritized-generative-replay-D3-004",
      "claim": "Compare PGR against REDQ augmented with prioritized experience replay (PER) using both TD-error and curiosity-based priority functions on state-based DMC tasks over 100K steps, measuring average return to determine whether densifying the replay distribution with conditional generation outperforms simply reweighting past experience via PER.",
      "source": "Section 5.1, Figure 3a"
    },
    {
      "id": "prioritized-generative-replay-D3-005",
      "claim": "Compare Curiosity-PGR against REDQ and SYNTHER each augmented with an intrinsic curiosity reward bonus (weight 0.1) on state-based DMC tasks (quadruped-walk, cheetah-run) over 100K steps over 3 seeds, measuring average return to determine whether PGR's benefits go beyond simply improving exploration.",
      "source": "Section 5.1, Figure 3b; Appendix B.1, Table 7"
    },
    {
      "id": "prioritized-generative-replay-D3-006",
      "claim": "Compare sample efficiency of Curiosity-PGR against SYNTHER and REDQ on state-based DMC tasks (cheetah-run, quadruped-walk, finger-turn-hard) by plotting learning curves of average return as a function of environment steps over 100K interactions, with 5 seeds, showing that Curiosity-PGR matches SYNTHER's final performance approximately 50K steps earlier.",
      "source": "Section 5.2, Figure 4a"
    },
    {
      "id": "prioritized-generative-replay-D3-007",
      "claim": "Compare sample efficiency of Curiosity-PGR against SYNTHER and DRQ-V2 on pixel-based DMC tasks (walker-walk, cheetah-run) by plotting learning curves of average return as a function of environment steps over 100K interactions, with 5 seeds, showing Curiosity-PGR consistently improves over DRQ-V2 while SYNTHER is eventually overtaken.",
      "source": "Section 5.2, Figure 4b"
    },
    {
      "id": "prioritized-generative-replay-D3-008",
      "claim": "Measure the faithfulness of generated transitions to true environment dynamics by computing mean-squared error (MSE) between generated and ground-truth next states and rewards for SYNTHER and Curiosity-PGR at epoch 50 over 10K generated transitions across 3 OpenAI Gym environments, testing whether conditional generation improves sample quality.",
      "source": "Section 5.2, Figure 5"
    },
    {
      "id": "prioritized-generative-replay-D3-009",
      "claim": "Quantify policy overfitting on the quadruped-walk task by measuring dormant ratio (DR, fraction of inactive neurons below an activation threshold) of policy networks over the course of training for REDQ, SYNTHER, and Curiosity-PGR, showing that Curiosity-PGR maintains a low and stable DR while REDQ exhibits high and increasing DR.",
      "source": "Section 5.2, Figure 6a"
    },
    {
      "id": "prioritized-generative-replay-D3-010",
      "claim": "Characterize how curiosity conditioning mitigates overfitting by measuring the distribution of curiosity F-values (Eq. 5) over 10K real transitions every 10K timesteps on quadruped-walk for Curiosity-PGR, and comparing against the same F-function evaluated on 10K real transitions from the unconditionally-trained SYNTHER baseline, demonstrating a growing left-skewed and long-tailed distribution for PGR.",
      "source": "Section 5.2, Figure 6b"
    },
    {
      "id": "prioritized-generative-replay-D3-011",
      "claim": "Evaluate how PGR and SYNTHER scale with increased policy network capacity by scaling hidden layers from 2x256 to 3x512 (approximately 6x more parameters) while increasing batch size from 256 to 1024 to maintain per-parameter throughput, on the quadruped-walk DMC task over 100K steps with 3 seeds, measuring average return.",
      "source": "Section 5.3, Figure 7a"
    },
    {
      "id": "prioritized-generative-replay-D3-012",
      "claim": "Evaluate how PGR and SYNTHER behave when varying the fraction of synthetic data per training batch (synthetic data ratio r) from 0.5 to 0.75 to 0.875 by doubling the batch size (256 to 512 to 1024) while keeping real transitions fixed at 128 per batch, on quadruped-walk over 100K steps with 3 seeds, measuring average return to determine how much synthetic data can replace real data.",
      "source": "Section 5.3, Figure 7b"
    },
    {
      "id": "prioritized-generative-replay-D3-013",
      "claim": "Combine larger policy networks (3x512), increased synthetic data ratio (r=0.75, batch 512), doubled UTD ratio (20 to 40), and doubled synthetic buffer capacity (1M to 2M) to evaluate whether PGR can leverage these components complementarily on quadruped-walk over 100K steps with 3 seeds, measuring average return.",
      "source": "Section 5.3, Figure 7c"
    },
    {
      "id": "prioritized-generative-replay-D3-014",
      "claim": "Validate that PGR is a framework-level contribution by conditioning on two alternative relevance functions, Random Network Distillation (RND) prediction error and Context-Tree Switching (CTS) pseudo-counts, evaluating PGR (RND) and PGR (CTS) against DRQ-V2, SYNTHER, and PGR (Curiosity) on pixel-based DMC tasks (walker-walk, cheetah-run) over 100K steps with 5 seeds, measuring average return.",
      "source": "Appendix A.1, Table 4"
    },
    {
      "id": "prioritized-generative-replay-D3-015",
      "claim": "Evaluate PGR in the randomized DMLab noisy-TV environment, a procedurally-generated 3D maze with stochastic visual observations (lower-right 42x42 quadrant replaced with per-pixel uniform noise) and sparse goal rewards (+10), comparing PGR conditioned on ICM, RND, and ECO relevance functions against PPO baselines with and without exploration bonuses on both Sparse and Very Sparse variants over 10M environment steps with 10 seeds, measuring average return.",
      "source": "Appendix A.2, Table 5"
    },
    {
      "id": "prioritized-generative-replay-D3-016",
      "claim": "Integrate PGR with implicit exploration methods by replacing the policy Q-networks with NoisyNets (learnable mu+sigma parameters with sampled epsilon noise per linear layer) and Bootstrapped DQN (ensemble of K Q-networks with per-episode Bernoulli(1) mask for updates), evaluating PGR (NoisyNets) and PGR (Boot-DQN) against standalone NoisyNets, Boot-DQN, and PGR (Curiosity) on state-based DMC tasks (quadruped-walk, cheetah-run) over 100K steps with 3 seeds, measuring average return.",
      "source": "Appendix B.2, Table 6"
    },
    {
      "id": "prioritized-generative-replay-D3-017",
      "claim": "Evaluate the robustness of PGR against model-based RL algorithms (MAX, DREAMER-V3) under noisy transition dynamics by adding isotropic Gaussian noise to 20% of transitions per batch (to the ICM forward dynamics for PGR, to the model ensemble for MAX, to the recurrent state-space world model for DREAMER-V3) on state-based DMC tasks (cheetah-run, walker-walk, hopper-hop) over 100K steps with 5 seeds, measuring average return under both original (clean) and noised conditions.",
      "source": "Appendix C, Table 8"
    }
  ],
  "D4": [
    {
      "id": "prioritized-generative-replay-D4-001",
      "claim": "Experiment phases: REDQ + Curiosity / SYNTHER + Curiosity: policy receives r_total = r_extrinsic + 0.1 * r_curiosity -> PGR (Curiosity): curiosity only used as conditioning signal for diffusion generation, not added to policy reward -> All methods evaluated on quadruped-walk and cheetah-run state-based DMC over 100K steps, 3 seeds -> PGR outperforms both baselines augmented with exploration bonuses.",
      "source": "Section 5.1 Fig. 3b; Appendix B.1 Table 7"
    },
    {
      "id": "prioritized-generative-replay-D4-002",
      "claim": "Experiment phases: Train REDQ, SYNTHER, Curiosity-PGR to 100K env steps on quadruped-walk -> Periodically compute DR by measuring fraction of neurons below activation threshold in policy networks -> Plot DR curves over training (Fig 6a) -> Curiosity-PGR shows low and stable DR; REDQ shows high and increasing DR, indicating aggressive overfitting.",
      "source": "Section 5.2 Fig. 6a"
    },
    {
      "id": "prioritized-generative-replay-D4-003",
      "claim": "Experiment phases: Train MAX, DREAMER-V3, PGR for 100K env steps on both original and noised conditions on cheetah-run, walker-walk, hopper-hop -> For noised condition: at each training step, add isotropic Gaussian noise to 20% of transitions -> For PGR: noise only affects ICM forward dynamics loss computation; diffusion model still generates clean transitions -> For MAX: noise affects model ensemble training data -> For DREAMER-V3: noise affects world model state inputs -> PGR demonstrates greater robustness with smaller performance degradation under noise.",
      "source": "Appendix C Table 8"
    }
  ]
}