Datasets:

ArXiv:
File size: 7,307 Bytes
3e10edb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
"""

Plotting churn ratio difference between partial and full observability

"""

import jax.numpy as jnp
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
import seaborn as sns
from jax import lax
from scipy.interpolate import interp1d

import wandb


def f(name):
    WINDOW_SIZE = 100
    SIGMA = 100
    INTERP_POINTS = 1000
    NORMALIZING_FACTOR = 200

    ENV_MAX_STEPS = {
        "CountRecallEasy": 2e7,
        "CountRecallMedium": 2e7,
        "CountRecallHard": 2e7,
        "BattleShipEasy": 2e7,
        "BattleShipMedium": 2e7,
        "BattleShipHard": 2e7,
        # other environments with default max steps 1e7
    }
    AXIS_FONT = {"fontsize": 9, "labelpad": 8}
    TICK_FONT = {"labelsize": 8}

    api = wandb.Api()
    runs = api.runs("bolt-um/Arcade-RLC-Churn")
    filtered_runs = [run for run in runs if run.state == "finished"]
    print(f"Total runs: {len(runs)}, Completed runs: {len(filtered_runs)}")

    METRIC_MAPPING = {
        "PQN": {"churn_ratio": "churn_ratio", "time_col": "env_step"},
        "PQN_RNN": {"churn_ratio": "churn_ratio", "time_col": "env_step"},
        "default": {"churn_ratio": "churn_ratio", "time_col": "TOTAL_TIMESTEPS"},
    }

    def process_run(run):
        """Process individual W&B run with dynamic max steps per environment"""
        try:
            config = {k: v for k, v in run.config.items() if not k.startswith("_")}
            env_name = config.get("ENV_NAME", "UnknownEnv")
            partial_status = str(config.get("PARTIAL", False))

            if env_name in ENV_MAX_STEPS:
                env_max_step = ENV_MAX_STEPS[env_name]
            else:
                env_max_step = 1e7

            alg_name = config.get("ALG_NAME", "").upper()
            memory_type = "MLP"
            if alg_name == "PQN_RNN":
                memory_type = config.get("MEMORY_TYPE", "Unknown").capitalize()

            metric_map = METRIC_MAPPING.get(alg_name, METRIC_MAPPING["default"])
            history = list(
                run.scan_history(
                    keys=[metric_map["churn_ratio"], metric_map["time_col"]]
                )
            )
            history = pd.DataFrame(
                history, columns=[metric_map["churn_ratio"], metric_map["time_col"]]
            )

            history["true_steps"] = history[metric_map["time_col"]].clip(
                upper=env_max_step
            )
            history = history.sort_values(metric_map["time_col"]).drop_duplicates(
                subset=["true_steps"]
            )

            if len(history) < 2:
                print(f"Skipping {run.name} due to insufficient data points")
                return None

            # Get first and last values for extrapolation
            first_return = history[metric_map["churn_ratio"]].iloc[0]
            last_return = history[metric_map["churn_ratio"]].iloc[-1]

            # Create unified interpolation grid for this environment
            unified_steps = np.linspace(0, env_max_step, INTERP_POINTS)
            unified_steps = np.round(unified_steps, decimals=5)
            scale_factor = NORMALIZING_FACTOR / env_max_step

            # Interpolate returns to uniform grid
            interp_func = interp1d(
                history["true_steps"],
                history[metric_map["churn_ratio"]],
                kind="linear",
                bounds_error=False,
                fill_value=(first_return, last_return),
            )
            interpolated_churn_ratio = interp_func(unified_steps)

            return pd.DataFrame(
                {
                    "Algorithm": f"{alg_name} ({memory_type})",
                    "churn_ratio": interpolated_churn_ratio,
                    # "Smoothed Return": smoothed_returns,
                    # "Cummax Return": np.array(cummax_returns),  # Convert back to NumPy
                    "True Steps": unified_steps,
                    "EnvName": env_name,
                    "Partial": partial_status,
                    "Seed": str(config.get("SEED", 0)),
                    "run_id": run.id,
                    "StepsNormalized": unified_steps / env_max_step,
                    "EnvMaxStep": env_max_step,
                    "ScaleFactor": scale_factor,
                }
            )

        except Exception as e:
            print(f"Error processing {run.name}: {str(e)}")
        return None

    # Process all runs and combine data
    # all_data = [df for run in filtered_runs if (df := process_run(run)) is not None]

    # if not all_data:
    #     print("No valid data to process")
    #     exit()
    # runs_df = pd.concat(all_data, ignore_index=True)
    # runs_df.to_pickle("churnratiodata.pkl")

    runs_df = pd.read_pickle("churnratiodata.pkl")
    # print(f"Total runs processed: {runs_df}")

    diff_df = pd.DataFrame()

    for env_name in runs_df["EnvName"].unique():
        env_data = runs_df[runs_df["EnvName"] == env_name]

        partial_true = env_data[env_data["Partial"] == "True"]
        partial_false = env_data[env_data["Partial"] == "False"]

        merged = pd.merge(
            partial_true[["StepsNormalized", "churn_ratio"]],
            partial_false[["StepsNormalized", "churn_ratio"]],
            on="StepsNormalized",
            suffixes=("_true", "_false"),
            how="inner",
        )

        merged["churn_diff"] = np.abs(
            merged["churn_ratio_true"] - merged["churn_ratio_false"]
        )
        merged["EnvName"] = env_name.replace("Easy", "")

        # diff_df = pd.concat([diff_df, merged[['EnvName', 'StepsNormalized', 'churn_diff']]], ignore_index=True)
        merged["churn_diff_cummax"] = merged.groupby("EnvName")["churn_diff"].cummax()
        # merged['churn_diff_avg'] = merged.groupby('EnvName')['churn_diff'].transform('mean')
        merged["churn_diff_avg"] = merged.groupby("EnvName")["churn_diff"].transform(
            lambda x: x.rolling(window=20, min_periods=1).mean()
        )

        diff_df = pd.concat(
            [
                diff_df,
                merged[
                    [
                        "EnvName",
                        "StepsNormalized",
                        "churn_diff",
                        "churn_diff_cummax",
                        "churn_diff_avg",
                    ]
                ],
            ],
            ignore_index=True,
        )

    plt.figure(figsize=(12, 7))
    sns.set()
    sns.lineplot(
        data=diff_df,
        x="StepsNormalized",
        y="churn_diff_avg",
        hue="EnvName",
        palette="Spectral",
        linewidth=2.5,
    )

    plt.title("Relative Policy Churn", fontsize=35)
    plt.xlabel("Training Progress", fontsize=35)
    plt.ylabel("POMDP/MDP Difference", fontsize=35)
    plt.tick_params(axis="both", which="major", labelsize=35)
    plt.legend(title="", loc="upper left", fontsize=20, ncol=2)
    plt.grid(True, alpha=0.5)
    plt.tight_layout()
    plt.savefig("{}.pdf".format(name), dpi=300, bbox_inches="tight", facecolor="white")


for i in range(1):
    f(f"churn{i}")
    print(f"churn{i} done")