Documentation Index
Fetch the complete documentation index at: https://mintlify.com/MilesONerd/neurenix/llms.txt
Use this file to discover all available pages before exploring further.
Overview
The Agent class provides a foundation for implementing various types of agents, such as reinforcement learning agents or autonomous agents.
Class Definition
class Agent:
def __init__(self, name: Optional[str] = None)
Parameters
The name of the agent. If None, a random name is generated.
Properties
name
@property
def name(self) -> str
Core Methods
act
def act(self, observation: Any) -> Any
The current observation of the environment.
learn
def learn(self, experience: Any) -> None
The experience to learn from.
reset
Reset the agent’s state. This is typically called at the beginning of a new episode.
Persistence
save
def save(self, path: str) -> None
The path to save the agent to.
load
def load(self, path: str) -> None
The path to load the agent from.
Example Usage
import neurenix as nx
from neurenix.agent import Agent
# Define a custom agent
class MyAgent(Agent):
def __init__(self, name=None):
super().__init__(name)
self.policy = None # Your policy network
def act(self, observation):
# Implement action selection logic
if self.policy is None:
return random.choice([0, 1, 2, 3]) # Random action
# Use policy network
with nx.Tensor.no_grad():
action = self.policy(observation)
return action
def learn(self, experience):
# Implement learning logic
state, action, reward, next_state, done = experience
# Update policy based on experience
loss = self.compute_loss(state, action, reward, next_state, done)
loss.backward()
self.optimizer.step()
def reset(self):
super().reset()
# Reset any episode-specific state
self._episode_rewards = []
# Create an agent
agent = MyAgent(name="MyCustomAgent")
# Use the agent in an environment
observation = env.reset()
for step in range(1000):
# Agent selects an action
action = agent.act(observation)
# Environment steps
next_observation, reward, done, info = env.step(action)
# Agent learns from experience
experience = (observation, action, reward, next_observation, done)
agent.learn(experience)
observation = next_observation
if done:
agent.reset()
observation = env.reset()
# Save the trained agent
agent.save("my_agent.pth")
# Load the agent later
new_agent = MyAgent()
new_agent.load("my_agent.pth")
Built-in Agent Types
Neurenix provides several built-in agent implementations:
DQN Agent
Deep Q-Network agent for discrete action spaces
A2C Agent
Advantage Actor-Critic agent
PPO Agent
Proximal Policy Optimization agent
SAC Agent
Soft Actor-Critic agent for continuous control