chore: merge on each trace

Author: qew21

rdagent/scenarios/data_science/proposal/exp_gen/merge.py

@@ -176,34 +176,34 @@ def gen(
                 )
             )
 
-        success_fb_list = list(set(trace_fbs))
-        logger.info(
-            f"Merge Hypothesis: select {len(success_fb_list)} from {len(trace_fbs)} SOTA experiments found in {len(leaves)} traces"
-        )
+        trace_leaf_summaries = []
+        for leaf_id, leaf in enumerate(leaves):
+            if leaf == trace.current_selection[0]:
+                continue
 
-        if len(success_fb_list) > 0:
-            exp_to_merge_fb_desc = T("scenarios.data_science.proposal.exp_gen.merge:trace").r(
-                exp_and_feedback_list=success_fb_list,
-                type="success",
-                heading="Successful iterations:",
-                success_trial_desc="These trials are the steps or changes that led to the success of the solution to be merged",
-                pipeline=DS_RD_SETTING.coder_on_whole_pipeline,
+            exp_fbs = trace.experiment_and_feedback_list_after_init(
+                return_type="sota",
+                search_type="ancestors",
+                selection=(leaf,),
+                max_retrieve_num=max_sota_retrieved_num_per_trace,
             )
-        else:
-            exp_index = self.get_exp_index(trace)
-            exp_to_merge_fb = trace.sota_experiment_fb(selection=(exp_index,))
-            if exp_to_merge_fb is None:
-                exp_to_merge_fb = trace.hist[exp_index]
-
-            exp_to_merge_fb_desc = T("scenarios.data_science.share:describe.feedback").r(
-                exp_and_feedback=exp_to_merge_fb,
-                heading="The feedback for the solution to be merged",
+            trace_leaf_summaries.append(
+                {
+                    "leaf_id": leaf_id,
+                    "experiments": exp_fbs,
+                    "best_exp": exp_fbs[-1][0],
+                }
             )
 
+        merge_fb_desc = T("scenarios.data_science.proposal.exp_gen.merge:merge_trace").r(
+            trace_leaf_summaries=trace_leaf_summaries,
+            success_trial_desc="These trials are the steps or changes that led to the success of the solution to be merged",
+        )
+
         component_desc = T("scenarios.data_science.share:component_description_in_pipeline").r()
         hypothesis_dict = self.hypothesis_gen(
             component_desc=component_desc,
-            exp_feedback_list_desc=exp_to_merge_fb_desc,
+            exp_feedback_list_desc=merge_fb_desc,
             sota_exp_desc=sota_exp_desc,
             enable_idea_pool=DS_RD_SETTING.enable_knowledge_base,
             pipeline=DS_RD_SETTING.coder_on_whole_pipeline,

rdagent/scenarios/data_science/proposal/exp_gen/merge.yaml

@@ -42,6 +42,39 @@ trace: |-
     {% endif %}
     {% endfor %}
     {% endif %}
+merge_trace: |-
+    {% if not trace_leaf_summaries or trace_leaf_summaries|length == 0 %}
+    No trace leaves {% if type == "success" %}with SOTA{% elif type == "failure" %}containing failed{% endif %} experiments available.
+    {% else %}
+    {% for leaf in trace_leaf_summaries %}
+    # Trace Leaf: {{ leaf.leaf_id }}
+    ### Recent Experiments (max 4)
+    {% if not leaf.experiments or leaf.experiments|length == 0 %}
+    No eligible experiments found on this leaf.
+    {% else %}
+    {% for exp_and_feedback in leaf.experiments %}
+    ## Experiment {{ loop.index }}
+    Target Problem: {{ exp_and_feedback[0].hypothesis.problem_desc }}
+    Proposed Hypothesis: {{ exp_and_feedback[0].hypothesis.hypothesis }}
+    Surpass Previous SOTA: {{ exp_and_feedback[1].decision }}
+    {% if exp_and_feedback[0].result is none %}
+    Experiment Score: Running buggy
+    Experiment Error: {{ exp_and_feedback[1].reason }}
+    {% else %}
+    Experiment Score: {{ exp_and_feedback[0].result.loc["ensemble"].iloc[0] }}
+    Experiment Feedback: {{ exp_and_feedback[1].reason }}
+    {% if exp_and_feedback[1].code_change_summary %}Code Change Summary: {{ exp_and_feedback[1].code_change_summary }}{% endif %}
+    {% endif %}
+    {% endfor %}
+    {% if leaf.best_exp %}
+    ### Best Code (last experiment)
+    {{ leaf.best_exp.experiment_workspace.all_codes }}
+    {% else %}
+    No final code available for this leaf.
+    {% endif %}
+    {% endif %}
+    {% endfor %}
+    {% endif %}
 hypothesis_gen:
   system: |-
     {% include "scenarios.data_science.share:scen.role" %}
@@ -66,7 +99,11 @@ hypothesis_gen:
     2. Note on Time/Memory Constraints
       - If prior experiments failed due to time/memory limitations, assume your new hypothesis will face the same constraints. In this case, prioritize efficiency and **ONLY** response to the problems related to time/memory constraints in your response dictionary.
       - Besides, do not compromise performance merely for efficiency since the current SOTA implementation do not encounter the constraints. You should think about how to balance the efficiency and performance so that your new hypothesis can be executed successfully and achieve satisfactory performance. 
-
+    3. Multi-Branch Learning
+      - When multiple branches (e.g., tabular, image, ensemble, workflow) have been explored, extract stable wins (changes that improved results repeatedly) and fragile wins (improvements with high variance or heavy cost).
+      - Prefer single-variable changes; if a proposal touches multiple components, split it into multiple hypotheses.
+      - When porting ideas from other branches, state the minimal integration and the specific acceptance/rollback condition.
+      
     # Task 2: Hypothesis Evaluation
     ## Evaluation Instruction
     Firstly, you should tag the hypothesis with one of the following components. If the hypothesis is related to multiple components, you should choose the most relevant one.

rdagent/scenarios/data_science/proposal/exp_gen/router/__init__.py

@@ -77,23 +77,10 @@ async def async_gen(self, trace: DSTrace, loop: LoopBase) -> DSExperiment:
             if loop.get_unfinished_loop_cnt(loop.loop_idx) < RD_AGENT_SETTINGS.get_max_parallel():
                 # set trace current selection
                 leaves: list[int] = trace.get_leaves()
-                if not timer.started or timer.remain_time() >= timedelta(hours=DS_RD_SETTING.merge_hours):
-                    local_selection = await self.trace_scheduler.next(trace)
+                local_selection = await self.trace_scheduler.next(trace)
 
-                    # set the local selection as the global current selection for the trace
-                    trace.set_current_selection(local_selection)
-                else:
-                    if len(leaves) < 2:
-                        local_selection = (-1,)
-                        trace.set_current_selection(selection=local_selection)
-                    else:
-                        local_selection = (leaves[0],)
-                        if trace.sota_exp_to_submit is not None:
-                            for i in range(1, len(leaves)):
-                                if trace.is_parent(trace.exp2idx(trace.sota_exp_to_submit), leaves[i]):
-                                    local_selection = (leaves[i],)
-                                    break
-                        trace.set_current_selection(local_selection)
+                # set the local selection as the global current selection for the trace
+                trace.set_current_selection(local_selection)
 
                 ds_plan = self.planner.plan(trace) if DS_RD_SETTING.enable_planner else DSExperimentPlan()