Branch and bound#22
Open
AdamJovine wants to merge 60 commits into
Open
Conversation
- Add Groq LLM class to llms.py and wire into load_model.py - Add hard constraint budget pre-filtering to nesy_agent - Simplify TPCAgent to delegate fully to LLMDrivenAgent - Update eval_tpc.py and pyproject.toml dependencies - Exclude .env, zips, __MACOSX, and egg-info from git Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
… venues in NesyAgent - Add Ollama class to llms.py for local model inference (strips <think> tags) - Wire ollama/ollama:<model> into init_llm with OLLAMA_BASE_URL/OLLAMA_MODEL env vars - load_dotenv at startup in run_tpc.py so .env is picked up automatically - Pre-filter restaurants and attractions in nesy_agent.py that cannot be reached before their closing time (using 20-min travel lower bound) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Points listen to a0ba55f which fixes stage2 dropping activity_position timing-constrained restaurants from their meal slots and changes the stage1/stage2 tie-breaking to prefer stage1. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
All listen development now tracks the china branch of LISTEN-active. Includes timing-venue pinning fix, --rerun-failures flag, constraint diagnostics, and inner-city budget improvements. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
_run_tournament_topk returns (winners, run_logs) but all call sites in solve() were assigning the raw tuple, making transport_winners etc. a 2-tuple of lists instead of a list of dicts. The fallback loop would AttributeError immediately on dict access. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…tes/agent Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
… budget, inner-city repair Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
… B&B solver Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…erminus) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…e-by fix Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
… slots Brings in 3 commits from LISTEN-active china branch (a1deb3c → e45d291): - Fix mixed-mode transport: go/back types parsed independently - Port Pass 2 apostrophe fix into constraints.py - Fix MISSING_REQUIRED_RESTAURANT: add meal slots + B&B skeleton sorting Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…SSING_REQUIRED_RESTAURANT Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
… pools Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…get-trim repair Picks up two fixes for ATTRACTION_COST constraint failures: - _prefilter_by_budget now runs before type-pinning so free-only queries don't get paid attractions injected via required-type pins - New budget-trim repair move in B&B for tight non-zero budgets - full_snippet used for required-type detection in repair moves Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
… guard Picks up two more ATTRACTION_COST fixes: - B&B repair now detects attraction_budget failures via label/full_snippet (abbreviated snippet was missing the <=N part, blocking budget-trim) - Assembler guards Day 1 lunch against calling next_restaurant past window, preventing required restaurants from being permanently consumed Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
… constraint Merges LISTEN-active fixes: - assembler.py: early-exit next_restaurant() when meal window already closed - bnb_agent.py: GT pinning safety net, repair injection via computed diagnostics, updated _MISSING_RNAME_RE to match &, NL2SL fallback to GT constraints - agent.py + bnb_agent.py: min_beds hotel filtering from new NL2SL field - constraints.py: accommodation_numbed() evaluator function - nl2sl.py: min_beds schema field, renderer, and few-shot example Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…hing Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…rmalization + cuisine aliases + nl2sl fixes Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
… + misc fixes Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Bumps listen from 6dfbe0a to 3410603 - Includes: hotel pin sort fix, poi_timing apostrophe regex fix, word-set name matching, whitespace normalization, nl2sl caching, rank cache atomic writes Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…ming fix Fixes chronological order violation where transport arrival time exceeds next activity start time (~16/18 affected plans in BNB_21 run). Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…ol fix Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
… fix eval score formula listen → 32ec728: - bnb_agent: _dedup_plan_attractions() strips repeated attractions from final plan output without changing BnB search paths (0 regressions). - agent: _rows_for_pins() skips DB entries already seen to handle nl2sl splitting compound venue names into multiple pins for the same DB row. eval_tpc.py: fix final_score formula — was using micro_comm twice (typo); now correctly uses macro_comm for the second 0.1 weight. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- build_nl2sl_corpus.py: extracts (NL, GT) pairs from run directories or zips, reverse-maps DSL programs to JSON params, writes rank_cache/nl2sl_corpus.json for use with --nl2sl-corpus flag in run_tpc.py - listen submodule: nl2sl dynamic few-shot + GT reverse-mapper + assembler fixes (Fix 8: Day 1 hotel reachability probe; Fix 9: restaurant canonical name) Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…+ expand attraction pool (DAV fix) Points listen to commit 8761432 which adds three DAV-improving assembler changes. Diagnosed from LISTEN_v2 evaluation: plans averaged only 1.34 attractions/day vs the 4.0/day target, yielding DAV=33%. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…rough cpsat DAV was near-zero (4.43%) because: 1. MTZ subtour elimination had wrong sign (+BigM → -BigM) making it trivial 2. CP-SAT selected attractions were being dropped by the routing MIP Both bugs fixed; expected DAV improvement ~4% → ~25%. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
… submodule - eval_submission.py: local replication of TPC@IJCAI2026 weighted-sum evaluator - show_failures.py, smart_merge.py, merge_rerun.py, fix_transport_price.py: analysis/repair utilities - test_nl2sl_models.py: NL2SL model comparison harness - nl2sl_corpus_v4y.json, v4_results.json: NL2SL training corpus and eval results - AGENT_README.md, BNB_EXPERIMENT_LOG.md, diagrams.html: experiment documentation - FULL1000_* / cpsat_*.log: full-1000 run logs for BNB and CP-SAT solvers - failing_queries/: cached query JSONs for persistent failures (for debugging) - .gitignore: exclude rank_cache/, .claude/, and LISTEN archive dirs - listen: update submodule to include hotel proximity filter + pair-distance objective Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…kfast Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…ration Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…ntroid Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…r/agent/bnb fixes Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…tive terms Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Fixes ATTRACTION_VISIT_TIME failures where spanning time constraints (start <= T1 AND end >= T2) caused the routing MIP to drop venues. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
assembler: hotel-return budget probes now use metro mode (conservative upper bound) instead of cheapest, preventing inner_city_budget overruns caused by walk/metro threshold races near 2 km. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
… formula fix Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Arrival (Day 1) and departure (Day D) now each get their own pool of _N_TRAVEL_DAY_REST=8 attractions instead of sharing 4. Fixes structural DAV ceilings caused by Day D always getting 0 attractions. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…es + cpsat IC fixes Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…ttr types Bumps LISTEN to ba14bbe which extends the CP-SAT IP to directly select restaurants and attractions on arrival/departure travel days, fixing: - Forced restaurants (required-name) on 2-day trips where n_full=0 - Required attraction types on 2-day trips - Global at-most-once uniqueness now spans full days + travel days Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…erify mode Add LISTEN_v3 run outputs (66 JSON files), challenging/easy UID lists, run_remaining.py script, and updated log files for cpsat ceiling analysis. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
…ed int cast - listen: assembler transit-day early hotel, cross-pool dedup, breakfast overhaul, colgen algorithm, nl2sl room_type rescue, cpsat meal window fix - nesy_agent.py: cast hotel_sel["numbed"] to int to avoid type errors - tpc_agent/utils.py: re-export shared UrbanTrip utilities for tpc_agent use Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.