From bcf842f9128fe98f74a9956783b827fec04c26d9 Mon Sep 17 00:00:00 2001
From: WinstonLiyt <104308117+WinstonLiyt@users.noreply.github.com>
Date: Tue, 24 Sep 2024 18:42:14 +0800
Subject: [PATCH] change the mechanism of traceback and action choosing (#320)

---
 rdagent/scenarios/kaggle/proposal/proposal.py | 6 +++---
 rdagent/utils/workflow.py                     | 2 +-
 2 files changed, 4 insertions(+), 4 deletions(-)

diff --git a/rdagent/scenarios/kaggle/proposal/proposal.py b/rdagent/scenarios/kaggle/proposal/proposal.py
index 9efd3c0f..a8e1f06e 100644
--- a/rdagent/scenarios/kaggle/proposal/proposal.py
+++ b/rdagent/scenarios/kaggle/proposal/proposal.py
@@ -91,8 +91,8 @@ class KGHypothesisGen(ModelHypothesisGen):
         self.reward_estimates = {
             "Feature engineering": 0.0,
             "Feature processing": 0.0,
-            "Model feature selection": 0.0,
-            "Model tuning": 0.5,
+            "Model feature selection": 0.2,
+            "Model tuning": 1.0,
         }
         self.confidence_parameter = 1.0
         self.initial_performance = 0.0
@@ -192,7 +192,7 @@ class KGHypothesisGen(ModelHypothesisGen):
             else:
                 performance_t_minus_1 = self.initial_performance
 
-            reward = performance_t_minus_1 - performance_t
+            reward = (performance_t - performance_t_minus_1) / performance_t_minus_1
             n_o = self.action_counts[last_action]
             mu_o = self.reward_estimates[last_action]
             self.reward_estimates[last_action] += (reward - mu_o) / n_o
diff --git a/rdagent/utils/workflow.py b/rdagent/utils/workflow.py
index b49b6af6..0c1f369b 100644
--- a/rdagent/utils/workflow.py
+++ b/rdagent/utils/workflow.py
@@ -117,7 +117,7 @@ class LoopBase:
                     continue
                 except CoderError as e:
                     logger.warning(f"Traceback loop {li} due to {e}")
-                    self.step_idx -= 1
+                    self.step_idx = 0
                     continue
 
                 end = datetime.datetime.now(datetime.timezone.utc)