chore: properly developing

engine/lib/behavior.py

@@ -0,0 +1,48 @@
+from sim.rl.behavior_loader.models import BehaviorModel, AgentBehaviorModel, aggregate_event_transitions
+import pandas as pd
+import numpy as np
+from .demand import generate_demand
+
+base_dir = "/home/velocitatem/Documents/Projects/PHANTOM/experiments"
+human_dir, agent_dir = f"{base_dir}/collected_data/", f"{base_dir}/agents/collected_data/"
+
+
+def adjust_behavior_to_condition(condition, transition_matrix):
+    # transition matrix just maps probability of eventA to eventB
+    # we enhance this that eventA-productI to eventB-productJ... based on the condition of interest
+    # this is to simulate the impact of demand onto the behavior
+    # NxN -> (N*(P + 1))x(N*(P + 1)) where P is number of products
+    new_transitions = transition_matrix.copy()
+    for col in new_transitions.columns:
+        for product in range(len(condition)):
+            # adjust the transition probability based on the demand condition
+            newname = f"{col}_product{product}"
+            new_transitions[newname] = new_transitions[col] * (condition[product] / np.sum(condition))
+    for row in transition_matrix.index:
+        for product in range(len(condition)):
+            newname = f"{row}_product{product}"
+            new_transitions.loc[newname] = new_transitions.loc[row] * (condition[product] / np.sum(condition))
+
+    return new_transitions.fillna(0.0)
+
+def sample_behavior(condition, human=True, max_len=40):
+    model = BehaviorModel(human_dir) if human else AgentBehaviorModel(agent_dir)
+    mdp = model.build_MDP()
+    raw_events = aggregate_event_transitions(mdp) # this gets us transtition between events (blind to products or prices)
+    # staet: {state: p} is raw_events we needc a matrix a pivot table
+    events_pivot = pd.DataFrame(raw_events).fillna(0.0)
+    # now adjust the transition matrix based on the condition to get a more informed transition matrix
+    adjusted_transitions = adjust_behavior_to_condition(condition, events_pivot)
+
+    trajectory = [np.random.choice(adjusted_transitions.index)]
+    while len(trajectory) < max_len or 'checkout' in trajectory[-1]:
+        probs = adjusted_transitions.loc[trajectory[-1]].values
+        sample = np.random.choice(adjusted_transitions.columns, p=probs/np.sum(probs) if np.sum(probs) > 0 else None)
+        trajectory.append(sample)
+    return trajectory
+
+if __name__ == "__main__":
+    t=sample_behavior(generate_demand(np.array([10,20,30])), human=True)
+    print(t)
+    t=sample_behavior(generate_demand(np.array([10,20,30])), human=False)
+    print(t)

engine/lib/demand.py

@@ -0,0 +1,40 @@
+import numpy as np
+
+def generate_demand(prices):
+    # assumption 1: each product has an intrinsic valuation drawn from a normal distribution centered at 50
+    product_valuations = np.random.normal(loc=50.0, scale=10.0, size=len(prices))
+    # assumption 2: demand decreases as price increases, following a simple linear model
+    demand = np.maximum(0, product_valuations - prices) # demand cannot be negative
+    demand = demand / np.sum(demand) * 100  # normalize to total demand of 1000 units so demand output is within [0, 100]
+    return demand
+
+def estimate_demand(trajectories):
+    demand_estimate = {}
+    for traj in trajectories:
+        for event in traj:
+            if 'view_product' in event:
+                product_id = int(event.split('_')[-1].replace('product', ''))
+                demand_estimate[product_id] = demand_estimate.get(product_id, 0) + 1
+    total_views = sum(demand_estimate.values())
+    for product_id in demand_estimate:
+        demand_estimate[product_id] = (demand_estimate[product_id] / total_views) * 100  # normalize to percentage
+    return demand_estimate
+
+# Example usage
+if __name__ == "__main__":
+    np.random.seed(42)
+    prices = np.array([20.0, 35.0, 50.0, 65.0])
+    demand = generate_demand(prices)
+    print("Generated Demand:", demand)
+    from .behavior import sample_behavior
+    N, alphat =200, 0.1
+    trajectories = []
+    for _ in range(int(N*(1 - alphat))):
+        trajectories.append(sample_behavior(demand, human=True))
+    for _ in range(int(N*alphat)):
+        trajectories.append(sample_behavior(demand, human=False))
+    demand_estimate = estimate_demand(trajectories)
+    print("Estimated Demand from Behavior:", demand_estimate)
+    delta = {k: demand_estimate.get(k, 0) - demand[i] for i, k in enumerate(range(len(prices)))}
+    delta = np.mean([np.abs(v) for v in delta.values()])
+    print("Demand Delta:", delta)

sim/rl/behavior_loader/models.py

@@ -226,6 +226,7 @@ if __name__ == "__main__":
 
     agent_model = AgentBehaviorModel(agent_dir)
     agent_mdp = agent_model.build_MDP()
+    print(agent_mdp)
     print(f"AGENT... Built MDP: {agent_mdp['num_states']} states, "
           f"{sum(len(t) for t in agent_mdp['transitions'].values())} transitions")
     if not agent_mdp['states']:
@@ -234,6 +235,9 @@ if __name__ == "__main__":
 
     human_evt = aggregate_event_transitions(human_mdp)
     agent_evt = aggregate_event_transitions(agent_mdp)
+    print(agent_evt)
+
+
     common = set(human_evt.keys()) & set(agent_evt.keys())
 
     if not common: