Skip to content

explanation.md

Latest commit

 

History

History
59 lines (45 loc) · 3.34 KB

explanation.md

File metadata and controls

59 lines (45 loc) · 3.34 KB

Parser Agent Explanation

This document explains the architecture and workflow of the current agent implementation in the parser repository.

Overview

The agent is designed to perform tasks in a web browser environment using Monte Carlo Tree Search (MCTS) for planning. It moves away from pixel-based planning to a DOM-based approach, using the capabilities of the browser's Document Object Model to understand and interact with the page.

Key Components

The codebase is modular, with each component handling a specific aspect of the agent's lifecycle:

1. main.py (Entry Point)

  • Role: Configures and launches the agent.
  • Functionality:
    • Instantiates specific implementations of ActionGenerator, RewardModel, and PriorPolicy.
    • Configures the MCTSPlanner with search parameters (simulations, depth, top-k).
    • Creates an AgentRunner and a MockBrowserEnv.
    • Runs a baseline episode and prints the results.

2. runner/ (Execution Loop)

  • AgentRunner: The main loop that orchestrates the agent-environment interaction.
    • Observes the current state.
    • Queries the planner for the next best action.
    • Executes the action in the environment.
    • Records the trace of events.
  • MockBrowserEnv: A placeholder environment that simulates a browser's state and responses for testing purposes without a real browser instance.

3. search/ (Planning Core)

  • MCTSPlanner: Implements the Monte Carlo Tree Search algorithm.
    • Uses ActionGenerator to expand the tree.
    • Uses PriorPolicy to guide selection.
    • Uses RewardModel to evaluate leaf nodes.
  • MCTSConfig: define hyperparameters like number of simulations (simulations=80), rollout depth (rollout_depth=5), etc.
  • PriorPolicy: A heuristic or learned model that suggests promising actions to explore first.

4. action_space/ (Action Generation)

  • ActionGenerator: Responsible for analyzing the current DOM state and enforcing rules to generate a list of valid, executable actions (e.g., click, type, scroll).

5. dom_encoder/ (Perception)

  • Encoder: Transforms the raw DOM tree into a structured state representation (embedding or feature vector) that the models and planner can process.

6. reward/ (Evaluation)

  • RewardModel: A component that assigns a scalar value to a state or action sequence, indicating how close the agent is to completing the user's objective.

7. traces/ (Debugging)

  • TraceRecorder: Captures the full history of the agent's reasoning process, including search trees, selected actions, and environment feedback, for later analysis.

Workflow

  1. Initialize: main.py builds the components and the runner.
  2. Observe: The runner asks the environment for the current state (DOM).
  3. Plan: The MCTSPlanner builds a search tree:
    • Select: Choose a promising node using Upper Confidence Bound (UCB).
    • Expand: Generate possible next actions using ActionGenerator.
    • Simulate: Estimate the value of the new state using the RewardModel or a rollout policy.
    • Backpropagate: Update the values of ancestor nodes.
  4. Act: The best action from the search is selected and executed in the real (or mock) environment.
  5. Repeat: The loop continues until the task is done or a limit is reached.