An Efficient Task-Oriented Dialogue Policy: Evolutionary Reinforcement Learning Injected by Elite Individuals

Ben Niu; Libo Qin; Shihan Wang; Yangyang Zhao

REVIEW 4 major objections 5 minor 10 references

An Efficient Task-Oriented Dialogue Policy: Evolutionary Reinforcement Learning Injected by Elite Individuals

T0 review · 4 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A hybrid of evolutionary search and deep Q-learning can train task-oriented dialogue policies more efficiently than DQN or LLM baselines.

desk verdict A plausible hybrid DRL+EA dialogue policy paper whose efficiency claim is undercut by an epoch-count comparison that hides a 4x difference in environment interactions. read the letter →

arxiv 2506.03519 v2 pith:GOWE2Y4X submitted 2025-06-04 cs.CL

Yangyang Zhao , Ben Niu , Libo Qin , Shihan Wang This is my paper

classification cs.CL

keywords task-orienteddialoguepolicylearningevolutionaryreinforcementeliteindividualinjectionexploration-exploitationtrade-offdeepQ-networkmulti-domain

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that a task-oriented dialogue policy trained by combining an evolutionary algorithm with a deep Q-network, plus an elite individual injection mechanism, balances exploration and exploitation better than either approach alone. The motivation is that DRL dialogue policies get trapped in local optima or waste samples, while standalone evolutionary search is too slow on natural-language tasks. EIERL reports higher success rates and rewards than DQN variants, noisy and curiosity-driven DQN agents, and prompt-based LLM policies at epochs 250 and 500 across three single-domain benchmarks and MultiWOZ2.1. If the mechanism holds, dialogue-policy training becomes more sample-efficient without expert demonstrations or high-quality user simulators.

What carries the argument

The load-bearing object is the elite individual injection (EII) mechanism, an adaptive threshold discriminator on cumulative reward. Each evaluated individual's fitness is compared with a running maximum $f_{\max}$; when the current best $f'_{\max}$ exceeds it, that elite policy's parameters replace the whole EA population and the threshold rises. This gives evolutionary search a gradient-free but guided direction, while the DQN population provides gradient-based local optimization and the shared replay buffer supplies experience. It is the component that turns slow population-only exploration into practical training for dialogue.

What would settle it

Train EIERL and the best DQN baseline under equal wall-clock time or equal total environment interactions instead of equal epochs; if the success-rate and reward gaps at epochs 250 and 500 shrink to noise, the efficiency claim is not supported.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central discovery is that adaptively injecting the best-performing evolved individual back into the evolutionary population, whenever its cumulative reward exceeds a running threshold, removes the main obstacle to using evolutionary reinforcement learning in dialogue. In the EIERL algorithm, a DRL agent is trained from a shared replay buffer while an EA population explores diverse policies; an elite discriminator tracks the best fitness seen so far and, on a new record, copies that individual's weights into every EA member instead of running another round of selection, crossover, and mutation. The paper reports that this injection accelerates convergence relative to plain ERL and that the full method outperforms DQN-based baselines and LLM policies on movie, restaurant, taxi, and MultiWOZ dialogue tasks. It also claims this is the first effective integration of EA and DRL for task-oriented dialogue policy.

Load-bearing premise

The comparison treats equal numbers of training epochs as equal training cost, although EIERL evaluates multiple evolved agents per epoch and stores only a fraction of their experience, so the reported efficiency may be an artifact of the epoch budget.

Editorial extensions

If this is right

Dialogue policy learning can become more sample-efficient: EIERL reaches high success rates by epoch 250, where plain DQN is still well below its final performance.
Evolutionary reinforcement learning, previously confined to game tasks, becomes applicable to natural-language dialogue domains.
The elite-injection schedule is adaptive rather than fixed, so it does not require a hand-tuned injection interval for each domain.
Because EIERL's DRL component is a standard Q-network, the method can combine with other off-policy RL algorithms or with richer fitness criteria.
On multi-domain dialogues, the same exploration-exploitation balance carries over, so EIERL offers a unified policy module for task-oriented dialogue systems.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The reported epoch-count advantage may not translate to wall-clock or compute parity: EIERL evaluates several agents per epoch and stores only a fraction of their experiences, so a cost-matched comparison is the real test of efficiency.
The adaptive threshold can be read as a simple form of success-history guidance; a natural test is whether tracking novelty or diversity of behaviors, rather than reward alone, improves it further.
The paper's single fitness criterion limits it to cumulative reward; a multi-criteria variant is an obvious extension and would be needed for complex domains where dialogue length, user satisfaction, and task completion conflict.
If elite injection generalizes, it could be applied to population-based training in other sequential decision tasks, not only dialogue, wherever evolution is slow and a cheap fitness signal exists.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Quick take: this is a plausible hybrid DRL+EA method for dialogue policy, with a simple elite injection trick, but the main efficiency claim rests on comparing equal epoch counts while EIERL runs four times as many environment interactions per epoch. The convergence-stage results may hold, but the efficiency argument needs rework.

What's actually new: applying evolutionary RL to task-oriented dialogue policy, plus the EII mechanism that replaces the EA population with the current best individual when a running max is beaten. That's simple but described clearly, and the experiments cover three single-domain simulators plus MultiWOZ. Code is claimed public, which helps.

What it does well: the paper is straightforward, the ablation showing EII helps vs plain ERL is useful, and the limitation paragraph is honest. The writing is clear.

Where it's soft: the stress-test note is correct. Algorithm 3 evaluates every individual in pop_evo ∪ pop_policy each epoch (m=3 plus n=1 in single-domain, so four dialogues per epoch versus one for DQN). The 1/M sampling only cuts replay buffer storage, not environment interactions or wall-clock. So 'higher success rates with fewer training epochs' is not supported by an epoch comparison; at epoch 50 EIERL is worse on Movie, and by epoch 500 it has consumed roughly 4x the experience. That's a real flaw in the central efficiency claim. The performance-at-convergence claim is separate and might be fine.

Also: Table 1 has a typo in Restaurant LLM_DP Turns going 20.16 to 29.16 with same success/reward; variances are deferred to Appendix D, not shown in main table; and hyperparameters (epsilon for baselines, P, sigma) are tuned on the same evaluation datasets, which is common but should be acknowledged. These are fixable.

Verdict: conditional accept. The method is clearly described and the empirical core is salvageable if the authors either equalize compute/environment interactions or reframe the claim as convergence-stage performance rather than efficiency, plus fix the reporting issues.

For you: if you work on dialogue policy or hybrid evolutionary/RL methods, worth reading. I'd bring it to reading group for the evaluation-methodology lesson. I wouldn't cite it in my own work, but it deserves a serious referee.

Referee Report

4 major / 5 minor

Summary. The paper proposes EIERL, a hybrid algorithm that combines deep reinforcement learning (DQN) with an evolutionary algorithm for task-oriented dialogue policy learning. The main novelty is an Elite Individual Injection (EII) mechanism that adaptively replaces the EA population with the best-performing individual when a fitness threshold is exceeded, with the goal of reducing the long evolution times that plague ERL in dialogue tasks. Experiments on three single-domain datasets (movie, restaurant, taxi) and MultiWOZ 2.1 compare EIERL against DQN baselines with different epsilon values, noisy DQN, curiosity-driven DQN, and LLM-based baselines. The paper reports higher success rates and rewards for EIERL at epochs 250 and 500 and claims that EIERL achieves a better exploration-exploitation balance and improved learning efficiency.

Significance. If the empirical claims are supported, the paper makes a useful contribution by being, to my knowledge, the first to integrate ERL with dialogue policy learning, and the EII mechanism is a plausible and well-motivated way to accelerate EA search. The public code release is a strength, and the multi-domain experiments are a welcome addition. No circularity issue is present: the fitness signal is the standard cumulative reward used in training, not an assumption of the conclusion. However, the central efficiency claim is currently not established, and several methodological gaps in the evaluation will need to be addressed before the results can be taken at face value.

major comments (4)

The efficiency claim, stated as 'higher success rates with fewer training epochs' in §4.2, is not supported because equal epoch counts do not correspond to equal training cost. In Algorithm 3, lines 9-15, every individual in pop_evo ∪ pop_policy is evaluated each epoch; for single-domain runs m=3 and n=1, so EIERL runs Evaluate four times per epoch, and each Evaluate performs ξ full dialogues. The DQN baselines run one dialogue per epoch. The 1/M sampling of stored experiences reduces replay-buffer storage, but it does not reduce environment interactions, fitness evaluation cost, or wall-clock time. Table 1 shows EIERL below DQN_EPSILON_0.05 at epoch 50 on Movie (0.2372 vs 0.3093) and no clear advantage on the other domains at that stage, so the reported advantage appears only after EIERL has consumed substantially more environment experience. The authors should either compare at matched total environment interactions / wall-clock time, or explicitly re-frame the claim as convergence-stage performance rather than efficiency.
The claim that 'all results of agent pairs are statistically significant at the same epoch (t-test, p < 0.05)' is unsupported by the manuscript. No variances, standard deviations, or confidence intervals are reported in Table 1, and Appendix D, titled 'Complete Results and Variances', contains only learning-curve figures, not variance tables or t-test details. Table 1 also contains a clear numerical inconsistency: for Rest. LLM_DP, the average turns at epoch 50 is 20.16, while at epochs 250 and 500 it is 29.16; since the LLM baseline is static and untrained, these numbers should be identical across epochs. This inconsistency undermines confidence in the table's accuracy and must be corrected.
The hyperparameters that most affect the method are selected on the same datasets used for final evaluation. Section 4.2 states that the optimal epsilon for DQN is chosen per domain, and §4.4.2 tunes the EA population size P and mutation strength σ on the same single-domain datasets, then uses those values as default settings for all experiments. This protocol does not control for overfitting to the test sets and weakens the claim that EIERL generalizes 'across various domains'. The authors should report results with hyperparameters selected on validation portions of the data, or at least provide a sensitivity analysis that separates tuning from final evaluation.
The choice of 500 epochs as the display cutoff is not justified by the stated convergence criterion. Appendix D asserts that 'at this point all methods have fully converged', but Table 1 shows EIERL's success rate still increasing substantially from epoch 250 (0.8033, Movie) to epoch 500 (0.8552), and similar increases in the Restaurant and Taxi domains. The cutoff therefore appears to favor the method under comparison rather than representing a common converged regime. The authors should report the full learning curves and justify the comparison point on the basis of convergence criteria that apply equally to all methods.

minor comments (5)

There is a typo in §3: 'exsiting experiences' should be 'existing experiences'.
The text says the effect of the EII mechanism is 'shown in Figure 4', but Figure 4 depicts the impact of mutation strength σ; the EII comparison appears in Figure 2. The cross-reference should be corrected.
Some table entries appear corrupted by formatting, such as the Rest. ICM_DQN row at epoch 500 ('0.0082 -32.88 9.25') and the Rest. LLM_DP turns inconsistency noted above. A careful proofread of the table is needed.
The LLM prompts in Appendix A are described as based on MultiWOZ 2.1, but the single-domain experiments use the Microsoft Dialogue Challenge datasets. It should be clarified whether the same prompts were used across all datasets and whether the prompt structure was adapted to the single-domain state representations.
The statement 'During the training phase of each epoch, the agents interact with the environment once' is ambiguous because EIERL has multiple agents. The sentence should specify whether this means one interaction per individual or one interaction per agent collectively; Algorithm 3 implies the former, which is the source of the cost-conflation issue in the major comments.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: EIERL is an empirical EA+DRL hybrid whose claimed gains are measured outcomes, not consequences of its own definitions.

full rationale

This is an empirical algorithm paper. The central claim in Section 4.3 that EIERL 'achieves superior performance across various domains due to a better balance between exploration and exploitation' is supported by measured success rates and rewards from environment interaction (Table 1, Figures 2-6), not derived from a fitted parameter or self-citation. The EII mechanism uses cumulative reward as fitness (Algorithm 2, Section 3.2.2), which is the same scalar reward used to train the DQN agent, but that is the task's objective function rather than a fitted input renamed as a prediction; the Limitation section explicitly acknowledges this single-criterion choice. Self-citations (Zhao et al. 2022, 2024) appear only as dataset/backbone references and are not load-bearing. The epoch-based efficiency comparison raises a fair-comparison concern, not circularity: EIERL evaluates m+n individuals per epoch while DQN baselines evaluate one, so 'fewer epochs' does not by itself imply lower computational cost, and hyperparameters P and sigma are selected on the test domains. These are experimental-validity risks outside the circularity definition. No quoted step reduces to its own input, so no circular step is reported and the score is 0.

Assumptions & free parameters 5 free parameters · 4 assumptions · 1 invented entities

The central claim rests on the empirical setup rather than on a derivation. The main free parameters are the EA population size and mutation strength, which are tuned on the evaluation datasets, plus several mutation probabilities and the fitness evaluation count that are not specified in the text. The axioms are simulator fidelity, the adequacy of cumulative reward as fitness, convergence by the chosen epoch cutoff, and computational cost parity across methods. The only invented entity is the elite injection mechanism, which has no falsifiable handle outside the reported experiments.

free parameters (5)

EA population size P = 3 (single-domain), 10 (multi-domain)
Tuned via experiments in Figure 3 on the same evaluation datasets; performance degrades with other values, so the reported gains depend on this choice.
Mutation strength sigma = 0.1
Tuned via experiments in Figure 4; step size per domain differs, and the default is selected after seeing results.
Elite ratio psi and mutation probabilities = Not specified in text (in code)
Introduced by Algorithm 1, but values are not listed in the paper; these affect the evolutionary dynamics and are not accounted for in the analysis.
DQN baseline epsilon = 0.05 per domain
Selected per domain based on Appendix C performance; the comparison baseline is tuned on the test domains.
Fitness evaluation episodes xi = Not specified
The number of dialogues used to estimate cumulative reward fitness is not reported, so fitness estimates may be noisy.

assumptions (4)

domain assumption The user simulator in the Microsoft Dialogue Challenge and ConvLab is a faithful proxy for real user behavior.
The experiments and conclusions about dialogue policy quality rely on simulator-based rewards and transitions; no human evaluation is reported.
domain assumption Cumulative reward is a sufficient fitness signal for selecting elite policies.
The EII mechanism and all EA selection use cumulative reward as fitness; the paper acknowledges in the Limitation section that a single criterion may not reflect true performance.
ad hoc to paper All compared methods have fully converged by 500 epochs (single-domain) and 10,000 epochs (multi-domain).
Appendix D states the 500-epoch cutoff is chosen because all methods have converged, but this is asserted rather than demonstrated with convergence criteria.
ad hoc to paper Sampling 1/M of experiences from M individuals makes computational cost comparable across methods.
Section 4.2 states this preserves consistent training costs, but it does not equalize environment interactions or evaluation cost per epoch.

invented entities (1)

Elite Individual Injection (EII) mechanism with elite discriminator
purpose: Periodically copies the best-performing individual's weights into all EA population members to accelerate evolution when the running maximum fitness threshold is exceeded.
This is a new algorithmic component introduced in Section 3.2.2. Its effectiveness is only demonstrated on the four datasets in this paper; it has no external falsifiable prediction outside these experiments.

how reviews work

0 comments

Cite this review

Pith. "Pith review of An Efficient Task-Oriented Dialogue Policy: Evolutionary Reinforcement Learning Injected by Elite Individuals." pith.science (2026). https://pith.science/paper/GOWE2Y4X

@misc{pith2026250603519,
  author       = {Pith},
  title        = {Pith review of: An Efficient Task-Oriented Dialogue Policy: Evolutionary Reinforcement Learning Injected by Elite Individuals},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GOWE2Y4X}},
  note         = {Machine review of arXiv:2506.03519}
}

read the original abstract

Deep Reinforcement Learning (DRL) is widely used in task-oriented dialogue systems to optimize dialogue policy, but it struggles to balance exploration and exploitation due to the high dimensionality of state and action spaces. This challenge often results in local optima or poor convergence. Evolutionary Algorithms (EAs) have been proven to effectively explore the solution space of neural networks by maintaining population diversity. Inspired by this, we innovatively combine the global search capabilities of EA with the local optimization of DRL to achieve a balance between exploration and exploitation. Nevertheless, the inherent flexibility of natural language in dialogue tasks complicates this direct integration, leading to prolonged evolutionary times. Thus, we further propose an elite individual injection mechanism to enhance EA's search efficiency by adaptively introducing best-performing individuals into the population. Experiments across four datasets show that our approach significantly improves the balance between exploration and exploitation, boosting performance. Moreover, the effectiveness of the EII mechanism in reducing exploration time has been demonstrated, achieving an efficient integration of EA and DRL on task-oriented dialogue policy tasks.

Figures

Figures reproduced from arXiv: 2506.03519 by the authors.

**Figure 2.** The effect of the EII mechanism on performance in Movie, Restaurant, and Taxi domains. [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

**Figure 3.** The effect of the EA population size (P) on performance across Movie, Restaurant, and Taxi domains. (a) Movie (b) Rest (c) Taxi [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Figures from the paper (5 more)

**Figure 4.** Figure 4: The impact of mutation strength (σ) on performance across Movie, Restaurant, and Taxi domains. method of encouraging exploration in new state spaces is not well-suited for task-oriented dialogue scenarios with clearly defined goals. While NOISY_DQN performs well in th…

**Figure 5.** Figure 5: Ablation experiment of two components (DRL and EA) of our method in Movie, Restaurant, and Taxi [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

**Figure 6.** Figure 6: Generalizability experiments of EIERL on MultiWOZ dataset with multiple domains. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

**Figure 7.** Figure 7: Effect of epsilon parameters on DQN performance [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]

**Figure 8.** Figure 8: The learning curves of different agents in Movie, Restaurant, and Taxi domains. [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

10 extracted references · 9 canonical work pages

[2]

Datasets and system roles: as the dialogue policy component in a task- oriented dialog system, you will make system decisions based on the Multi- WOZ 2.1 dataset

work page
[4]

ActionType

Generate system actions: based on the user dialog state, you need to generate system actions. These actions should be provided in the following format: [[“ActionType”, “Domain”, “Slot”, “Value”]] where ‘ActionType‘ denotes the type of action (e.g. Request, Inform, Confirm, etc.), ‘Domain‘ specifies the associated domain (e.g. restaurant, taxi, hotel, etc....

work page
[9]

Command execution requirements: when receiving a command, you must strictly follow the given instructions without performing any actions outside the scope of the command or generating any additional words

work page
[10]

Datasets and system roles: as the dialogue policy component and the natural language generation component of a task-based dialog system, you will make system decisions based on the MultiWOZ 2.1 dataset

work page
[11]

This state will be used as a basis for decision making

Processing user dialog state: you will receive a formatted user dialog state: Listing 1. This state will be used as a basis for decision making

work page
[12]

Inform

Generate system actions: generate your response to user conversations directly. Listing 1: Data Format 1 1{ 2’ user_action ’ : [ [ " Inform " , " Hotel " , " Area " , " east " ] , [ " Inform " , " Hotel " , " Stars " , " 4 " ] ] , 3’ system_action ’ : [ ] , 4’ belief_state ’ : { 5’ police ’ : { ’ book ’ : { ’ booked ’ : [ ] } , ’ semi ’ : { } } , 6’ hotel...

work page
[2013]

Rewarding What Matters: Step-by-Step Reinforcement Learning for Task-Oriented Dialogue

Exploration and exploitation in evolutionary al- gorithms: A survey.ACM Comput. Surv., 45(3):35:1– 35:33. Caibo Dong and Dazi Li. 2024. Adaptive evolutionary reinforcement learning with policy direction.Neural Processing Letters, 56(2):69. Huifang Du, Shuqin Li, Minghao Wu, Xuejing Feng, Yuan-Fang Li, and Haofen Wang. 2024. Reward- ing what matters: Step-...

work page Pith review arXiv 2024
[2019]

Deep Dyna-Q: Integrating Planning for Task-Completion Dialogue Policy Learning

Dialogue management with deep reinforce- ment learning: Balancing exploration and exploita- tion. In2019 8th Brazilian Conference on Intelligent Systems (BRACIS), pages 449–454. IEEE. Xuecheng Niu, Akinori Ito, and Takashi Nose. 2024. Scheduled curiosity-deep dyna-q: Efficient explo- ration for dialog policy learning.IEEE Access. Deepak Pathak, Pulkit Agr...

work page Pith review arXiv 2024

Show all 10 references

[2021]

Karush Suri

A survey on spoken language understanding: Recent advances and new frontiers.arXiv preprint arXiv:2103.03095. Karush Suri. 2022. Off-policy evolutionary reinforce- ment learning with maximum mutations. In21st International Conference on Autonomous Agents and Multiagent Systems...

2022 arXiv
[2024]

Zihao Yi, Jiarui Ouyang, Yuwen Liu, Tianhao Liao, Zhe Xu, and Ying Shen

Tree of thoughts: Deliberate problem solving with large language models.Advances in Neural Information Processing Systems, 36. Zihao Yi, Jiarui Ouyang, Yuwen Liu, Tianhao Liao, Zhe Xu, and Ying Shen. 2024. A survey on recent advances in llm-based multi-turn dialogue systems. C...

2024 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [2]

Datasets and system roles: as the dialogue policy component in a task- oriented dialog system, you will make system decisions based on the Multi- WOZ 2.1 dataset

work page

[2] [4]

ActionType

Generate system actions: based on the user dialog state, you need to generate system actions. These actions should be provided in the following format: [[“ActionType”, “Domain”, “Slot”, “Value”]] where ‘ActionType‘ denotes the type of action (e.g. Request, Inform, Confirm, etc.), ‘Domain‘ specifies the associated domain (e.g. restaurant, taxi, hotel, etc....

work page

[3] [9]

Command execution requirements: when receiving a command, you must strictly follow the given instructions without performing any actions outside the scope of the command or generating any additional words

work page

[4] [10]

Datasets and system roles: as the dialogue policy component and the natural language generation component of a task-based dialog system, you will make system decisions based on the MultiWOZ 2.1 dataset

work page

[5] [11]

This state will be used as a basis for decision making

Processing user dialog state: you will receive a formatted user dialog state: Listing 1. This state will be used as a basis for decision making

work page

[6] [12]

Inform

Generate system actions: generate your response to user conversations directly. Listing 1: Data Format 1 1{ 2’ user_action ’ : [ [ " Inform " , " Hotel " , " Area " , " east " ] , [ " Inform " , " Hotel " , " Stars " , " 4 " ] ] , 3’ system_action ’ : [ ] , 4’ belief_state ’ : { 5’ police ’ : { ’ book ’ : { ’ booked ’ : [ ] } , ’ semi ’ : { } } , 6’ hotel...

work page

[7] [2013]

Rewarding What Matters: Step-by-Step Reinforcement Learning for Task-Oriented Dialogue

Exploration and exploitation in evolutionary al- gorithms: A survey.ACM Comput. Surv., 45(3):35:1– 35:33. Caibo Dong and Dazi Li. 2024. Adaptive evolutionary reinforcement learning with policy direction.Neural Processing Letters, 56(2):69. Huifang Du, Shuqin Li, Minghao Wu, Xuejing Feng, Yuan-Fang Li, and Haofen Wang. 2024. Reward- ing what matters: Step-...

work page Pith review arXiv 2024

[8] [2019]

Deep Dyna-Q: Integrating Planning for Task-Completion Dialogue Policy Learning

Dialogue management with deep reinforce- ment learning: Balancing exploration and exploita- tion. In2019 8th Brazilian Conference on Intelligent Systems (BRACIS), pages 449–454. IEEE. Xuecheng Niu, Akinori Ito, and Takashi Nose. 2024. Scheduled curiosity-deep dyna-q: Efficient explo- ration for dialog policy learning.IEEE Access. Deepak Pathak, Pulkit Agr...

work page Pith review arXiv 2024