ACTIVE-SIMULATE-PLAN

This repo contains the source code for ACTIVE-SIMULATE-PLAN, a simulation-free implementation of Monte Carlo Tree Search (MCTS) for finding sequences of causal interventions that transform an initial symbolic state into a goal state for unknown anomalies.

Installation

ACTIVE-SIMULATE-PLAN has been tested on Ubuntu 22.04 LTS with Python 3.10. To install, run:

git clone --recurse-submodules https://github.com/convince-project/active-simulate-plan
cd active-simulate-plan/src/simulate-plan/downward
./build.py

Dependencies

ACTIVE-SIMULATE-PLAN uses Fast Downward for PDDL planning. This is included as a submodule and is handled by the above steps. Citation: Helmert, M. (2006). The Fast Downward Planning System. Journal of Artificial Intelligence Research, 26, 191-246.

ACTIVE-PLAN

ACTIVE-PLAN identifies anomalies by applying causal interventions with Monte Carlo Tree Search (MCTS) to run “what-if” analyses.

Architecture

src/active-plan/
└── identification/
    ├── mcts.py                # Core MCTS algorithm
    ├── state_manager.py       # Symbolic state tracking and updates
    ├── interventions.py       # Intervention parsing and application
    ├── reward_shaper.py       # Reward computation
   

config/
├── scenario1.json             # Block misalignment scenario
├── scenario2.json             # (to be added)
├── scenario3.json             # (to be added)
└── symbolic_state.json        # Initial world state

Quick Start

cd src/active-plan/identification
python3 mcts.py

When prompted, select a scenario (default is Scenario 1):

Select scenario (1-3, or press Enter for scenario1): 1

ACTIVE-PLAN will execute MCTS to find a sequence of interventions that achieves the goal state.

Running Examples

Upon running mcts.py, the system will:

1. Load the scenario configuration from config/scenario1.json
2. Load the initial symbolic state from config/symbolic_state.json
3. Execute MCTS search to identify corrective interventions
4. Save results to results_scenario1.json

SIMULATE-PLAN

SIMULATE-PLAN uses PDDL planning to generate recovery actions for geometrically misaligned blocks identified by MCTS.

Architecture

src/simulate-plan/
└── recovery/
    ├── planner.py           # PDDL recovery planner
    ├── pddl_builder.py      # PDDL problem generator 
    ├── plan_parser.py       # Fast Downward output parser
    └── domains/
        └── domain_s1.pddl   # PDDL domain for block stacking

Quick Start

cd src/simulate-plan/recovery
python3 planner.py

When prompted, select a scenario (default is Scenario 1):

Select scenario (1-3, or press Enter for scenario1): 1

Running Examples

Upon running planner.py, the system will:

1. Load the initial state from active-plan/config/symbolic_state.json
2. Load the goal state from active-plan/config/scenario1.json
3. Load MCTS interventions from active-plan/identification/results_scenario1.json
4. Generate PDDL problem with target predicates for misaligned blocks
5. Execute Fast Downward planner to find optimal action sequence
6. Parse and display the recovery plan
7. Save results to fd_output/plan_scenario1.json

API and Tutorials

The documentation for ACTIVE-SIMULATE-PLAN can be found here, where you can read more about the API and find tutorials.

Maintainer

This repository is maintained by:

Yazz Warsame	yazz	yazzwarsame@gmail.com