How driw0x/kaggriculture works
Kaggriculture develops autonomous farm-strategy agents for Kaggle. The main workflow is an observation from the game environment entering an agent, which returns market and worker actions for the environment to apply; the next turn supplies updated state. The repository also develops heuristic-agent families and a behavioral-cloning path from replay data through state/target encoding and model inference. The excerpts show the Jet agent’s observation-driven routing and policy decoding, but do not establish wiring among the broader agent variants or unsampled training modules.
Main entry points: policy.py BC policy · bc_dataset.py BC dataset · replay_dataset.py Replay dataset · jet6.py Jet agent
Components 17
Game interaction
- Kaggriculture environment
- Game observation
- Farm actions
Agent strategies
- Jet agent
src/agents/jet6.py - CHI planners
src/agents/chi.py - Ye agents
src/agents/ye.py - Agent interface
src/agents/agent.py - Kaggle submission
submissions/agent.py
Behavioral cloning
- BC agent
src/agents/bc_agent.py - BC policy
src/learning/policy.py - BC model
src/learning/model.py - State encoder
src/learning/state_encoder.py - Target encoder
src/learning/target_encoder.py
Replay learning
- Replay dataset
src/learning/replay_dataset.py - Decision extractor
src/learning/decision_extractor.py - BC dataset
src/learning/bc_dataset.py - Preencoded dataset
src/learning/preencoded_dataset.py
Connections 19
- Kaggriculture environment → Game observation: provides
- Game observation → Jet agent: supplies state
- Jet agent → Farm actions: returns actions
- Farm actions → Kaggriculture environment: applies turn
- Game observation → CHI planners: supplies state
- CHI planners → Farm actions: returns actions
- Game observation → Ye agents: supplies state
- Ye agents → Farm actions: returns actions
- Game observation → BC agent: supplies state
- BC agent → BC policy: requests action
- BC policy → State encoder: encodes state
- BC policy → BC model: runs inference
- BC policy → Target encoder: decodes output
- Replay dataset → BC dataset: provides examples
- Replay dataset → Decision extractor: extracts actions
- BC dataset → State encoder: encodes states
- BC dataset → Target encoder: encodes targets
- Preencoded dataset → BC dataset: provides encoded data
- Kaggle submission → Kaggriculture environment: submits agent