State vector design #4

New Issue

2026-08-15T22:05:18+02:00

SirStone commented

2026-08-15 22:05:18 +02:00

Question

What exactly goes into the observation vector fed to the actor-critic network?

Known inputs: own bot state (x, y, direction, speed, energy, gunDirection, gunHeat) + tracked enemy state (x, y, direction, speed, energy, age-of-last-scan). Normalize to [-1,1]?

Open sub-questions:

Include wall distances (4 directions)? Arena-relative position?
Include relative bearing/distance to enemy vs absolute positions?
Include own velocity components (vx, vy) vs scalar speed + heading?
Include bullet states (own in-flight bullets)?
Fixed vector size — what dimensions?

## Question What exactly goes into the observation vector fed to the actor-critic network? Known inputs: own bot state (x, y, direction, speed, energy, gunDirection, gunHeat) + tracked enemy state (x, y, direction, speed, energy, age-of-last-scan). Normalize to [-1,1]? Open sub-questions: - Include wall distances (4 directions)? Arena-relative position? - Include relative bearing/distance to enemy vs absolute positions? - Include own velocity components (vx, vy) vs scalar speed + heading? - Include bullet states (own in-flight bullets)? - Fixed vector size — what dimensions?

SirStone added the wayfinder:grilling label 2026-08-15 22:05:18 +02:00

SirStone referenced this issue

2026-08-15 22:05:50 +02:00

RL Bot — Wayfinder Map #1

SirStone commented

2026-08-15 22:47:55 +02:00

Resolution

State vector: 42 floats, all normalized to ~[-1, 1]

Current tick (14 floats):

#	Feature	Normalization
1-2	own x, y	/ arenaWidth, / arenaHeight
3	own direction	/ 360
4	own speed	/ 8
5	own energy	/ 100
6	own gunDirection	/ 360
7	own gunHeat	/ ~1.8
8-9	enemy x, y	/ arenaWidth, / arenaHeight
10	enemy direction	/ 360
11	enemy speed	/ 8
12	enemy energy	/ 100
13	hasFired	0 or 1
14	lastFirePower	/ 3

Derived (8 floats):

#	Feature	Normalization
15	enemy acceleration	/ 2
16	enemy turn rate	/ 10
17	relative bearing to enemy	/ 180 (signed)
18	distance to enemy	/ ~1700 (arena diagonal)
19-22	distance to 4 walls	/ max(arenaW, arenaH)

Enemy window — last 5 ticks (20 floats):

#	Feature	Normalization
23-42	enemy x, y, direction, speed (t-1 through t-5)	same as current tick

Design decisions

No own bot history — agent chose its own actions, current state captures momentum via speed.
Window size = 5 ticks — enough for "instinct" about enemy trajectory patterns.
RNN/GRU is the upgrade path if MLP+window hits a ceiling on pattern recognition.
Hand-crafted derived features (acceleration, turn rate, wall distances) complement the window — no need for the network to learn basic physics.

## Resolution ### State vector: 42 floats, all normalized to ~[-1, 1] **Current tick (14 floats):** | # | Feature | Normalization | |---|---------|--------------| | 1-2 | own x, y | / arenaWidth, / arenaHeight | | 3 | own direction | / 360 | | 4 | own speed | / 8 | | 5 | own energy | / 100 | | 6 | own gunDirection | / 360 | | 7 | own gunHeat | / ~1.8 | | 8-9 | enemy x, y | / arenaWidth, / arenaHeight | | 10 | enemy direction | / 360 | | 11 | enemy speed | / 8 | | 12 | enemy energy | / 100 | | 13 | hasFired | 0 or 1 | | 14 | lastFirePower | / 3 | **Derived (8 floats):** | # | Feature | Normalization | |---|---------|--------------| | 15 | enemy acceleration | / 2 | | 16 | enemy turn rate | / 10 | | 17 | relative bearing to enemy | / 180 (signed) | | 18 | distance to enemy | / ~1700 (arena diagonal) | | 19-22 | distance to 4 walls | / max(arenaW, arenaH) | **Enemy window — last 5 ticks (20 floats):** | # | Feature | Normalization | |---|---------|--------------| | 23-42 | enemy x, y, direction, speed (t-1 through t-5) | same as current tick | ### Design decisions - **No own bot history** — agent chose its own actions, current state captures momentum via speed. - **Window size = 5 ticks** — enough for "instinct" about enemy trajectory patterns. - **RNN/GRU is the upgrade path** if MLP+window hits a ceiling on pattern recognition. - **Hand-crafted derived features** (acceleration, turn rate, wall distances) complement the window — no need for the network to learn basic physics.

SirStone closed this issue

2026-08-15 22:47:56 +02:00

SirStone referenced this issue

2026-08-15 22:48:12 +02:00

RL Bot — Wayfinder Map #1

SirStone referenced this issue

2026-08-16 11:51:22 +02:00

Network architecture #9

SirStone referenced this issue

2026-08-16 12:19:20 +02:00

Network architecture #9

Sign in to join this conversation.