feat: draft - response rate scoring surface

- add SCORE_SCALING constant
- update response rate scoring logic

storb/constants.py

@@ -13,6 +13,9 @@
 PIECE_LENGTH_SCALING = 0.5
 PIECE_LENGTH_OFFSET = 8.39
 
+# Scaling factor for scoring
+SCORE_SCALING = 1
+
 MAX_UPLOAD_SIZE = 1 * 1024 * 1024 * 1024 * 1024  # 1 TiB
 
 # Error correction encoding parameters

storb/validator/reward.py

@@ -1,5 +1,7 @@
 import numpy as np
 
+from storb.constants import SCORE_SCALING
+
 
 def get_response_rate_scores(
     self,
@@ -28,9 +30,25 @@ def get_response_rate_scores(
         uids.append(uid)
         ret_attempts = max(miner_stats.get("retrieval_attempts", 1), 1)
         store_attempts = max(miner_stats.get("store_attempts", 1), 1)
-        retrieval_rate = abs(miner_stats.get("retrieval_successes", 0) / ret_attempts)
-        store_rate = abs(miner_stats.get("store_successes", 0) / store_attempts)
-        weighted_rate_sum = (retrieval_rate / 2) + (store_rate / 2)
+        retrieval_success_rate = abs(
+            miner_stats.get("retrieval_successes", 0) / ret_attempts
+        )
+        store_success_rate = abs(miner_stats.get("store_successes", 0) / store_attempts)
+
+        # TODO: do we want to "reset" the total attempts after a certain threshold?
+        # because at the moment for larger attempt values,
+        # the curve is very steep (small changes in success rate have a larger impact)
+        retrieval_rate_score = (
+            (1 / SCORE_SCALING) * np.log(ret_attempts) * np.log(retrieval_success_rate) + 1
+        )
+        store_rate_score = (
+            (1 / SCORE_SCALING) * np.log(store_attempts) * np.log(store_success_rate) + 1
+        )
+
+        retrieval_rate_score = np.clip(retrieval_rate_score, 0, 1)
+        store_rate_score = np.clip(store_rate_score, 0, 1)
+
+        weighted_rate_sum = (retrieval_rate_score / 2) + (store_rate_score / 2)
         weighted_rate_sums.append(weighted_rate_sum)
 
     uids = np.array(uids)