Attention Prioritized Experience Replay With Application to Self-Driving Cars: Learning From More Informative Experiences Helps Improve the Training Quality
Why this work is in the frame
A frame that forgets how it found something cannot be audited. These are the routes that admitted this work.
Bibliographic record
Abstract
Deep reinforcement learning (DRL) stands out as a powerful solution to the interconnected tasks involved in the control of self-driving cars. To perform control tasks rigorously, DRL models inevitably rely on high-quality observations which in turn make the training of such models computationally expensive. Improving the training efficiency of DRL models is therefore of critical importance, particularly for self-driving cars facing complicated driving scenes and scenarios. This paper proposes a novel attention-based targeted sampling method to improve training of deep Q-networks. Specifically, we utilize a convolutional neural network architecture with a multihead attention mechanism to enhance the agent’s emphasis on salient objects within the scene. The contribution of this work goes back to constructing a scoring mechanism based on the given level of attention to each agent’s experience. By using both the current and next states within each experience tuple, the constructed score encourages novelty in the agent’s training, which consequently speeds up the process. By dynamically prioritizing the replay of scenes using the attention score attributed by the agent, we propose an attention-based prioritized experience replay mechanism to expedite the agent’s training process. By visiting highly scored transitions more frequently, the agent learns from more informative experiences that helps with improving the training quality. By controlling the level of prioritization and bias correction for a driving scenario developed using the Carla simulator, the attained results verify the applicability and superiority of the proposed sampling scheme.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.000 | 0.000 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.000 | 0.000 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.000 | 0.000 |
| Open science | 0.000 | 0.000 |
| Research integrity | 0.000 | 0.000 |
| Insufficient payload (model declined to judge) | 0.000 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from it