shijazks/ad-placement-rl-agents

Web Ad-Placement RL Agents

Multiple RL/bandit models trained on a synthetic web ad-placement dataset: [link to your dataset repo].

Models included

File	Algorithm	Type
BC.pt	Behavior Cloning	Imitation learning baseline
CQL.pt	Conservative Q-Learning	Offline RL
FQI.joblib	Fitted Q-Iteration	Offline RL
LinUCB.joblib	LinUCB	Contextual bandit
surrogate_bc.joblib	Surrogate reward model for BC	Evaluation/simulation
surrogate_cql.joblib	Surrogate reward model for CQL	Evaluation/simulation
surrogate_fqi.joblib	Surrogate reward model for FQI	Evaluation/simulation
surrogate_linucb.joblib	Surrogate reward model for LinUCB	Evaluation/simulation
action_encoder.joblib	Action space encoder	Preprocessing
feature_encoders.joblib	Feature encoders	Preprocessing
feature_columns.joblib	Feature column schema	Preprocessing

Usage

import joblib
import torch

# example: loading LinUCB
linucb_model = joblib.load("LinUCB.joblib")

# example: loading a PyTorch model
bc_model = torch.load("BC.pt")

# load supporting encoders
feature_encoders = joblib.load("feature_encoders.joblib")
action_encoder = joblib.load("action_encoder.joblib")

Training data

Trained on synthetic data — see [dataset link] for details on generation and schema.

Limitations

Trained entirely on synthetic data; not validated on real user traffic or production ad-serving environments.