When Chaos Wins: Adding Noise Improved My Snake AI's Stability
The author discovered that turning off noise during evaluation of their Snake AI led to significantly worse and inconsistent performance compared to training. Contrary to standard practice, keeping the noise active improved stability and scores by preventing the AI from getting stuck in deterministic failure states. The noise, intended only for exploration during training, became a crucial component of the policy itself.
- ▪The author's Snake AI performed worse during deterministic evaluation when noise was turned off.
- ▪Noisy networks, designed to aid exploration during training, unexpectedly improved evaluation stability when left active.
- ▪Deterministic evaluation revealed a bimodal performance distribution, with 25% of runs failing completely.
- ▪With noise enabled, the AI avoided degenerate action loops and achieved higher average scores and lower variance.
- ▪The learned noise parameters (sigma weights) were found to be integral to the policy, not just training artifacts.
Opening excerpt (first ~120 words) tap to expand
try { if(localStorage) { let currentUser = localStorage.getItem('current_user'); if (currentUser) { currentUser = JSON.parse(currentUser); if (currentUser.id === 3896851) { document.getElementById('article-show-container').classList.add('current-user-is-article-author'); } } } } catch (e) { console.error(e); } Stat Phantom Posted on May 17 When Chaos Wins: Adding Noise Improved My Snake AI's Stability #machinelearning #deeplearning #ai #chaos Snake CNN: Rainbow DQN Exploration (4 Part Series) 1 A CNN Grid Encoding for Snake AI That DOUBLES! the Best Published Score 2 2 Lines of Code Saved 6.4x Memory on My Snake AI 3 Removing PER From Rainbow DQN Set a New Snake AI World Record 4 When Chaos Wins: Adding Noise Improved My Snake AI's Stability Greetings all! Continuing the series where I…
Excerpt limited to ~120 words for fair-use compliance. The full article is at DEV.to (Top).
Discussion
0 commentsMore from DEV.to (Top)
