π The First Self-Improving agents with RL / Prompting Optimization
Python agentaigptmulti-agentpythonreinforcement-learningself-improvementsociety-of-mindswarm-intelligence
#!/usr/bin/env python
# -*- coding: utf-8 -*-
from typing import Optional
from swarm.graph import Graph
from swarm.environment.operations.crosswords import Reflection, GreedySteps
from swarm.environment.agents.agent_registry import AgentRegistry
@AgentRegistry.register('CrosswordsReflection')
class CrosswordsReflection(Graph):
def __init__(self, domain: str, model_name: Optional[str] = None, meta_prompt: bool = False, num_reflections=1, num_inner_iters=2):
self.num_reflections = num_reflections
self.num_inner_iters = num_inner_iters
super().__init__(domain, model_name, meta_prompt)
def build_graph(self):
last_step = None
for i in range(self.num_reflections + 1):
for j in range(self.num_inner_iters):
step = GreedySteps(self.domain, self.model_name)
self.add_node(step)
if last_step:
last_step.add_successor(step)
last_step = step
else:
self.input_nodes = [step]
last_step = step
if i < self.num_reflections:
reflection = Reflection(self.domain, self.model_name)
self.add_node(reflection)
if last_step is None:
raise Exception("last_step should not be None")
last_step.add_successor(reflection)
last_step = reflection
if last_step is None:
raise Exception("last_step should not be None")
self.output_nodes = [last_step]