Update README.md

vwxyzjn · Feb 20, 2023 · 5f266c7 · 5f266c7
1 parent af0d1b4
commit 5f266c7
Showing 1 changed file with 2 additions and 1 deletion.
diff --git a/README.md b/README.md
@@ -19,6 +19,7 @@ Cleanba is CleanRL's implementation of DeepMind's Sebulba distributed training f
 
 **Scalable**: We can scale to N+ GPUs allowed by `jax.distributed` and memory (e.g., it can run with 16 GPUs). This makes cleanba suited for large-scale distributed training tasks such as RLHF.
 
+**Understandable**: We adopt the single-file implementation philosophy used in CleanRL, making our core codebase succinct and easy to understand. For example, our `cleanba/cleanba_ppo_envpool_impala_atari_wrapper.py` is ~800 lines of code.
 
 
 
@@ -139,4 +140,4 @@ To improve efficiency of Cleanba, we uses JAX and EnvPool, both of which are des
 
 [Espeholt et al., 2018](https://arxiv.org/abs/1802.01561) did not disclose the hardware usage and runtime for the Atari experiments. We did our best to recover its runtime by interpolating the results from the [R2D2 paper](https://openreview.net/pdf?id=r1lyTjAqYX) and found IMPALA (deep) takes ~2 hours.
 
-![](static/r2d2_impala.png)
+![](static/r2d2_impala.png)