Tower Defence

Overview

• Environment ID: kbediako/tower_defence
• Python package: prime_td_env
• Short description: Multi-turn macro-round tower defense environment for hosted RL.
• Tags: games, tower-defense, rl, verifiers, prime-rl, multi-turn
• Published package shape: env.py, pyproject.toml, README.md, src/

Datasets

• Primary dataset(s): Procedural tower-defense seeds generated by the environment.
• Snapshot mode: Configurable round snapshots (dataset.snapshots) for curriculum and stability studies.
• Split control: Training/eval sample volume is controlled by environment args and run config.

Task

• Type: Multi-turn game interaction (macro-round planning).
• Interaction contract: One assistant plan maps to one in-game round progression.
• Action surface: Candidate-index planning with {"type":"plan","actions":[{"type":"choose","index":N}, ...]}.

Quickstart

Run local smoke checks:

PYTHONPATH=src python3 scripts/smoke.py

Run baseline local evaluation:

PYTHONPATH=src python3 scripts/eval_baseline.py --episodes 10 --max-rounds 20 --output out/metrics.json

Run hosted training:

prime rl run configs/lab/prime-td.toml

Environment Arguments

Group	Key args	Purpose
wrapper	wrapper="macro_round"	Enables multi-turn round-by-round planning mode.
difficulty	max_rounds	Controls episode horizon/curriculum cap.
observation	max_action_candidates, max_build_slots, max_towers, max_threats	Bounds payload size and candidate space.
candidate_balance	min_build_frac, max_upgrade_candidates, by_phase.*	Tunes build/upgrade exposure by phase.
dataset	policy, rollout_steps, snapshots, safe_explore_*	Controls training observation generation.
rules	auto_advance_round, prep_actions_*, mask_sell	Governs turn semantics and allowed behavior.

Metrics

Metric	Meaning
reward/mean	Aggregate training reward over samples in a step.
metrics/num_turns	Episode length in environment turns.
format_reward	Validity of structured plan/action output.
macro_round_delta (derived)	Round advance per turn; expected delta_round == 1 after turn 1.
Action mix (derived)	Build vs upgrade distribution, tracked by phase.

Run Hygiene

• Record run IDs and config filenames immediately after launch.
• Pull rollouts for every sampled step and parse all user/assistant turns.
• Validate macro-round invariant delta_round == 1 for user observations after turn 1.
• If sample upload 500s appear, reduce tokens/observation caps or horizon pressure.

Detailed longitudinal results and run-by-run analysis: docs/RESULTS.md.