Tambayoyin Tattaunawar AI 6: Ka'idojin Hanyoyi Uku na AI Agent: ReAct, Plan-and-Solve da Reflection

Hanyoyi Uku na AI Agent: ReAct, Plan-and-Solve da Reflection

AI Agent wani mutum-mutumi ne mai iya fahimtar muhalli, yanke shawara, da aiwatar da ayyuka. Akwai manyan hanyoyi uku: ReAct, Plan-and-Solve, da Reflection. A kasa an bayyana su tare da zane-zane da misalan lamba.

1. ReAct (Reasoning + Acting)

Babban ra'ayi: Haɗa tunani (Reasoning) da aiki (Acting) a jere. Agent a kowane mataki yana tunanin halin yanzu da shirin gaba (tunani), sannan ya aiwatar da aiki (kamar kiran kayan aiki, bincike), sannan ya ci gaba da tunani bisa sakamakon.

Zane:

[Halin farko] → [Tunani: tunanin mataki na gaba] → [Aiki: aiwatar da aiki] → [Kallon sakamako] → [Tunani: sabunta shiri] → ... → [Amsa ta ƙarshe]

Misalin lamba (pseudo-code):

def react_agent(question):
    context = []
    while not solved:
        # Tunani: samar da matakin tunani
        thought = llm.generate_thought(question, context)
        # Aiki: zaɓi aiki bisa tunani
        action = llm.choose_action(thought)
        # Aiwatar da aiki, samun kallo
        observation = execute_action(action)
        # Ƙara tunani, aiki, kallo zuwa mahallin
        context.append((thought, action, observation))
    return final_answer

Misali:
- Mai amfani ya tambaya: "Yanayin Beijing yau yaya?"
- Agent tunani: "Ina buƙatar bincika API na yanayi, ina buƙatar sunan birni da kwanan wata."
- Aiki: kira API na yanayi (sigogi: Beijing, yau)
- Kallo: ya dawo "Rana, 25°C"
- Tunani: "Na samu bayani, zan iya amsawa."
- Fitarwa: "Beijing yau rana, 25°C."

2. Plan-and-Solve

Babban ra'ayi: Fara tsara cikakken shiri (Plan), sannan aiwatar da shi a hankali (Solve). A matakin shiri, an raba aiki mai rikitarwa zuwa matakai, a matakin aiwatarwa ana aiwatar da su a jere, ana iya daidaita shiri bisa sakamakon tsaka-tsaki.

Zane:

[Aiki] → [Tsara shiri: raba matakai] → [Aiwatar da mataki 1] → [Aiwatar da mataki 2] → ... → [Aiwatar da mataki N] → [Amsa ta ƙarshe]

Misalin lamba:

def plan_and_solve(task):
    # Matakin shiri
    plan = llm.generate_plan(task)  # misali: ["Bincike", "Tattara bayanai", "Rubuta rahoto"]
    context = {}
    for step in plan:
        # Aiwatar da kowane mataki
        result = execute_step(step, context)
        context[step] = result
    # Haɗa sakamako
    final = llm.synthesize(context)
    return final

Misali:
- Aiki: "Rubuta labari game da AI Agent"
- Shiri:
1. Bincika ma'anar AI Agent da sabbin ci gaba
2. Karanta da tattara mahimman bayanai
3. Rubuta tsarin labari
4. Cika abun ciki
5. Gyara da bugawa
- Aiwatarwa: aiwatar da kowane mataki a jere, a ƙarshe fitar da labari.

3. Reflection

Babban ra'ayi: Agent yana yin tunani (Reflection) game da ayyukansa yayin aiwatarwa ko bayan aiwatarwa, yana kimanta sakamako da inganta ayyuka na gaba. Yawanci ya haɗa da sukar kai, gyara kuskure, ko inganta dabara.

Zane:

[Aiki] → [Kallon sakamako] → [Tunani: kimanta ko nasara] → [Idan bai yi nasara ba: daidaita dabara] → [Aiki sake] → ... → [Nasara]

Misalin lamba:

def reflection_agent(task):
    max_attempts = 3
    for attempt in range(max_attempts):
        action = llm.generate_action(task)
        result = execute(action)
        # Tunani
        reflection = llm.reflect(task, action, result)
        if reflection['success']:
            return result
        else:
            # Daidaita bayanin aiki ko dabara bisa tunani
            task = reflection['improved_task']
    return None

Misali:
- Aiki: "Lissafta 1234 * 5678"
- Aiki: lissafta kai tsaye, sami sakamako 7006652
- Tunani: duba tsarin lissafi, gano kuskuren ɗauka
- Daidaita: sake lissafta, sami sakamako daidai 7006652 (daidai)
- Idan har yanzu kuskure, ci gaba da tunani har sai daidai.

Taƙaitawa da kwatanta

A aikace, ana haɗa su uku, misali ReAct tare da tsarin tunani, ko Plan-and-Solve inda kowane mataki yana tunani bayan aiwatarwa.