| from __future__ import annotations |
|
|
| import itertools |
|
|
| import numpy as np |
|
|
| from .config import BenchmarkConfig |
|
|
|
|
| def decode_continuous_action(raw_action: np.ndarray, config: BenchmarkConfig) -> np.ndarray: |
| market = config.market |
| action = np.nan_to_num(np.asarray(raw_action, dtype=np.float32), nan=0.0, posinf=1.0, neginf=0.0) |
| action = np.clip(action, 0.0, 1.0) |
| reward_ratio = market.min_reward_ratio + action[0] * (market.max_reward_ratio - market.min_reward_ratio) |
| liquidity_ratio = market.min_liquidity_ratio + action[1] * (market.max_liquidity_ratio - market.min_liquidity_ratio) |
| burn_rate = action[2] * market.max_burn_rate |
| return sanitize_actual_action(np.array([reward_ratio, liquidity_ratio, burn_rate], dtype=np.float32), config) |
|
|
|
|
| def encode_actual_action(actual_action: np.ndarray, config: BenchmarkConfig) -> np.ndarray: |
| market = config.market |
| reward, liquidity, burn = sanitize_actual_action(actual_action, config) |
| return np.array( |
| [ |
| (reward - market.min_reward_ratio) / (market.max_reward_ratio - market.min_reward_ratio), |
| (liquidity - market.min_liquidity_ratio) / (market.max_liquidity_ratio - market.min_liquidity_ratio), |
| burn / market.max_burn_rate, |
| ], |
| dtype=np.float32, |
| ).clip(0.0, 1.0) |
|
|
|
|
| def sanitize_actual_action(actual_action: np.ndarray, config: BenchmarkConfig) -> np.ndarray: |
| market = config.market |
| reward_ratio = float(np.clip(actual_action[0], market.min_reward_ratio, market.max_reward_ratio)) |
| liquidity_ratio = float(np.clip(actual_action[1], market.min_liquidity_ratio, market.max_liquidity_ratio)) |
| burn_rate = float(np.clip(actual_action[2], 0.0, market.max_burn_rate)) |
| total = reward_ratio + liquidity_ratio |
| if total > market.max_total_allocation: |
| scale = market.max_total_allocation / total |
| reward_ratio *= scale |
| liquidity_ratio *= scale |
| return np.array([reward_ratio, liquidity_ratio, burn_rate], dtype=np.float32) |
|
|
|
|
| def discrete_action_grid(config: BenchmarkConfig) -> np.ndarray: |
| values = np.linspace(0.0, 1.0, int(config.discrete_levels), dtype=np.float32) |
| return np.array(list(itertools.product(values, values, values)), dtype=np.float32) |
|
|
|
|
| def decode_discrete_action(action_index: int, config: BenchmarkConfig) -> np.ndarray: |
| grid = discrete_action_grid(config) |
| index = int(np.clip(action_index, 0, len(grid) - 1)) |
| return decode_continuous_action(grid[index], config) |
|
|
|
|
