INTERMEDIATEAI Arena
Advanced Model Guide — Buckets, Settings, Data Flow
Advanced model guide for View All: collect clean data, set Buckets, tweak Epochs/Batch Size/Learning Rate, use Remove Sparsity, Oversampling, and Multi-Stream.
By Esportz.fun Team
Archive · original publication date unavailable
10 min read
TL;DR
- Use the Advanced model with Exponential Memory Updating enabled so zone edits don’t wreck unrelated data.
- Record tight, isolated demonstrations: cut idle frames and avoid extra moves to keep training signals clean.
- Process first runs with high Epochs and Learning Rate, low Batch Size; enable Remove Sparsity and Oversampling early.
- Pick Buckets per ground/air zone and combine Discrete Distance + Your Direction + Angle to Opponent for reliable behavior.
What You Need to Know First
Prerequisites
| Item | Effect | Where to Get |
|---|---|---|
| Advanced model | Zone-isolated edits, advanced settings for converting bad data | Select from model menu before training |
| Exponential Memory Updating | Prioritizes recent training sessions for faster behavior shifts | Advanced model checkbox |
| Remove Sparsity | Deletes empty info cells to make training denser and faster | Processing options (95% usage) |
| Oversampling | Helps memorization of positions/directions; useful in initial training | Processing options |
| Multi-Stream | Boosts confirmed existing actions/directions proportionally | Processing options (use with care) |
| Buckets (On Ground / In Air) | Isolates training by situation (prevents cross-zone side effects) | Buckets selector in Advanced UI |
| Epochs, Batch Size, Learning Rate, Direction and Action Lambda | Core training knobs for behavior shaping | Processing tutorial / advanced settings |
Levels & Time
- No character levels are required, but expect multiple iterations. Early sessions establish baseline behavior; iteration is the meta here. Plan for several short record/process cycles rather than one long experiment.
Quick notes
- Ixtlanian’s tip: keep demos tight and surgical. A single sloppy recording can pollute a bucket. Think like a surgeon, not a drunk barber.
- The Advanced model is the top-tier pick for competitive players because you can edit "Side Of Stage" behavior without blowing up "On Ground" routines.
Step-by-Step Guide
-
Choose the Advanced model and enable Exponential Memory Updating
- Before you press record, switch to Advanced. Check Exponential Memory Updating for sessions where recent behavior matters more than legacy data.
-
Plan your session and pick Buckets
- Decide which Buckets you’ll record into. The isolation is the Advanced model’s main power.
| Item | Effect | Where to Get |
|---|---|---|
| On Ground: Projectile Active | Trained when special abilities are used on ground | On Ground bucket selector |
| On Ground: Opponent Stunned | Triggers when opponent is stunned | On Ground bucket selector |
| On Ground: Opponent Knockback | Triggers on opponent knockback events | On Ground bucket selector |
| On Ground: Opponent Active | Catches all other ground cases | On Ground bucket selector |
| In Air: Near Blast Zone / Safe Zone / Under Stage / Side Of Stage | Air zones trained independently | In Air bucket selector |
-
Collect data the right way
- Minimize Empty Information Slots: press the data button and act immediately. Long pauses add empty cells that disguise the signal.
- Avoid Unnecessary Actions: keep each demo isolated. If you want an uppercut after a dash, demo that sequence cleanly and nothing else.
-
Process initial training runs (establish baseline)
- Settings table for first runs
| Setting | Recommendation | Purpose |
|---|---|---|
| Epochs | High | Force the model to learn core patterns |
| Learning Rate | High | Allow meaningful parameter updates early |
| Batch Size | Low | Increase weight of scarce examples |
| Direction and Action Lambda | Low (initial) | Permit changes to actions/directions |
| Remove Sparsity | Enabled | Collapse empty cells into meaningful samples |
| Oversampling | Enabled (early) | Help model memorize positions/directions |
| Multi-Stream | Off initially | Avoid unintended biasing of other actions |
-
Harden or protect existing data
- If you want to adjust a small behavior without upsetting everything, increase Direction and Action Lambda to maximum and drop Epochs/Learning Rate. That tells the system: "don’t overwrite the good stuff." Think of it as putting a guardrail on your training.
-
Use Focus Settings to teach context-aware reactions
- Your Actions: good for sewing multi-step combos; use late in tuning.
- Opponent Actions: almost always on for ground training to react to enemy moves.
- Your Positioning / Opponent Positioning: use when spatial relations matter. Your Positioning is essential for edge recovery and platform returns.
- Your Direction + Angle to Opponent + Discrete Distance: classic combo for behavior that depends on facing and distance. This combo is your bread-and-butter for chase/grab/uppercut windows.
- Recoveries Remaining: only for "Side Of Stage" and "Near Blast Zone" to avoid repeating failed returns.
-
Iterate: don’t save bad sessions
- If a processed run produces garbage, don’t keep it. Figure out which focus or setting caused the regression and re-run with corrected demos. The model rewards a clean, surgical approach more than volume.
Pro Tips
- When teaching close-air mixups, record separate sessions for "Near Blast Zone" and "Safe Zone" so the model doesn’t conflate two different recovery priorities. Use Oversampling to reinforce near-edge behaviors.
- If a new action corrupts unrelated behavior, flip Direction and Action Lambda to max and rerun with low Epochs/Learning Rate to nudge only the new action.
- Use Multi-Stream only after you’ve tested changes in isolation. It accelerates confirming existing actions but can quietly weaken others if abused.
- For edge defense, combine Opponent Off Stage + Discrete Distance + Your Direction. It forces the fighter to treat off-stage enemies differently than regular ground pressure.
- Treat Remove Sparsity as the default for processing. You’ll waste less time debugging empty cells and training will converge cleaner and faster.
Common Mistakes to Avoid
- Recording long demos with idle frames — fills the bucket with empty info and muddies the signal.
- Using Multi-Stream or high Learning Rate without isolating Buckets — changes bleed into unrelated behaviors.
- Toggling many focus settings at once — if something breaks, you won't know which setting caused it.