---
license: cc-by-4.0
tags: [finance, defi, amm, reinforcement-learning]
pretty_name: RL Execution Training Logs
---

# RL Execution Training Logs

Per-checkpoint DQN training logs for
[arXiv:2607.10960](https://arxiv.org/abs/2607.10960) — *Reinforcement Learning
for Execution under Dynamic Fees in a Closed-Loop DEX Simulator*.

Primary artifacts (model checkpoints, result CSVs, code) live in the public
repo [github.com/egpivo/amm-lab](https://github.com/egpivo/amm-lab)
(`data/rl_equilibrium/`), commit `656ae56`. These training logs are
supplementary reproducibility evidence — validation implementation-shortfall
(IS) curves per episode, with the selected ("best") checkpoint marked — kept
separately since they were briefly excluded by an overly broad `.gitignore`
rule.

## Contents

`dqn_train_logs.tar.gz` — 6 logs, one per DQN variant:

- `dqn_train.log` (dynamic duopoly, the base variant)
- `dqn_train_completion_aware.log`
- `dqn_train_constant_duopoly.log`
- `dqn_train_dynamic_monopoly.log`
- `dqn_train_order_after.log`
- `dqn_train_order_random.log`

Each line: `ep <episode> eps <exploration rate> val IS <bps> (<elapsed s>)`,
with `<- best` marking the checkpoint that was selected and shipped.

## Note on reproducibility

The paper's reproducibility claim rests on the shipped `.pt` checkpoints
being exactly re-evaluable (content-hashed in
`data/rl_equilibrium/m3r_run_manifest.json`), not on these training curves.
These logs are included for transparency, not as a correctness dependency.