gqt
63 lines
1.7 KiB
Python
63 lines
1.7 KiB
Python
#!/usr/bin/env python3
|
|
# -*- coding: UTF-8 -*-
|
|
###########################################################################
|
|
# Copyright © 1998 - 2026 Tencent. All Rights Reserved.
|
|
###########################################################################
|
|
"""
|
|
Author: Tencent AI Arena Authors
|
|
|
|
Configuration for Robot Vacuum PPO agent.
|
|
清扫大作战 PPO 配置。
|
|
"""
|
|
|
|
|
|
class Config:
|
|
|
|
# Feature dimensions: 21x21x6 local map + scalar planning features + last action.
|
|
# 特征维度：21x21x6 多通道局部地图 + 标量规划特征 + 上一步动作。
|
|
VIEW_SIZE = 21
|
|
MAP_CHANNELS = 6
|
|
FEATURES = [
|
|
VIEW_SIZE * VIEW_SIZE * MAP_CHANNELS,
|
|
66, # global memory, charger, NPC, and action-improvement features
|
|
8, # last action one-hot / 上一步动作 one-hot
|
|
]
|
|
FEATURE_SPLIT_SHAPE = FEATURES
|
|
FEATURE_LEN = sum(FEATURES)
|
|
DIM_OF_OBSERVATION = FEATURE_LEN
|
|
|
|
# Action space: 8 directional moves
|
|
# 动作空间：8个方向移动
|
|
ACTION_NUM = 8
|
|
|
|
# Single-head value
|
|
# 单头价值
|
|
VALUE_NUM = 1
|
|
|
|
# PPO hyperparameters
|
|
# PPO 超参数
|
|
GAMMA = 0.99
|
|
LAMDA = 0.95
|
|
|
|
INIT_LEARNING_RATE_START = 0.00025
|
|
BETA_START = 0.008
|
|
BETA_END = 0.002
|
|
BETA_DECAY_STEPS = 4000
|
|
CLIP_PARAM = 0.2
|
|
VF_COEF = 0.5
|
|
PPO_EPOCHS = 3
|
|
MINI_BATCH_SIZE = 256
|
|
NORMALIZE_ADVANTAGE = True
|
|
TARGET_KL = 0.04
|
|
|
|
# Evaluation tie-break: when policy probabilities are close, prefer safer
|
|
# coverage/recharge actions with a lightweight heuristic.
|
|
EVAL_TIE_BREAK_PROB_GAP = 0.015
|
|
EVAL_TIE_BREAK_SCORE_SCALE = 0.01
|
|
|
|
LABEL_SIZE_LIST = [ACTION_NUM]
|
|
LEGAL_ACTION_SIZE_LIST = LABEL_SIZE_LIST.copy()
|
|
|
|
USE_GRAD_CLIP = True
|
|
GRAD_CLIP_RANGE = 0.5
|