Merge pull request #44 from leonvanbokhorst/Refactoring-generic-agent,-testing,-NER-implementation

Refactor generic agent and NER implementation

Author: leonvanbokhorst

requirements.txt

@@ -4,7 +4,7 @@ fonttools==4.54.1
 iniconfig==2.0.0
 kiwisolver==1.4.7
 matplotlib==3.9.2
-numpy==2.1.2
+numpy==2.0.2
 packaging==24.1
 pillow==10.4.0
 pluggy==1.5.0
@@ -17,3 +17,6 @@ six==1.16.0
 pydantic
 torch
 torchvision
+textblob
+spacy==3.8.2
+https://github.com/explosion/spacy-models/releases/download/en_core_web_sm-3.8.0/en_core_web_sm-3.8.0-py3-none-any.whl

src/active_inference_forager/agents/generic_agent.py

@@ -6,6 +6,9 @@
 from pydantic import Field, ConfigDict
 from collections import deque
 import random
+import re
+from textblob import TextBlob
+import spacy
 
 from active_inference_forager.agents.base_agent import BaseAgent
 from active_inference_forager.agents.belief_node import BeliefNode
@@ -80,9 +83,13 @@ class GenericAgent(BaseAgent):
     episode_lengths: List[int] = Field(default_factory=list)
     total_steps: int = Field(default=0)
 
+    # NLP model
+    nlp: spacy.language.Language = Field(default_factory=lambda: spacy.load("en_core_web_sm"))
+
     model_config = ConfigDict(arbitrary_types_allowed=True)
 
-    def __init__(self, state_dim: int, action_dim: int, **kwargs):
+    def __init__(self, action_dim: int, **kwargs):
+        state_dim = 17  # Updated to match the environment's state dimension
         super().__init__(state_dim=state_dim, action_dim=action_dim, **kwargs)
 
         self.root_belief = BeliefNode(
@@ -94,12 +101,19 @@ def __init__(self, state_dim: int, action_dim: int, **kwargs):
         self.target_network.load_state_dict(self.q_network.state_dict())
         self.optimizer = optim.Adam(self.q_network.parameters(), lr=self.learning_rate)
 
-        self.initialize_belief_and_action_space()
+        self.action_space = [
+            "ask_question",
+            "provide_information",
+            "clarify",
+            "suggest_action",
+            "express_empathy",
+            "end_conversation",
+        ]
         self.exploration_rate = self.epsilon_start
 
     def take_action(self, state: np.ndarray) -> str:
         if np.random.rand() < self.exploration_rate:
-            return np.random.choice(self.action_space)
+            return random.choice(self.action_space)
         state_tensor = torch.FloatTensor(state).unsqueeze(0).to(self.device)
         q_values = self.q_network(state_tensor)
         return self.action_space[q_values.argmax().item()]
@@ -274,19 +288,57 @@ def _build_belief_hierarchy(self, node: BeliefNode, level: int):
             self._build_belief_hierarchy(child, level + 1)
 
     def process_user_input(self, user_input: str) -> np.ndarray:
-        features = np.zeros(5)
+        features = np.zeros(17)  # Updated to match the environment's state dimension
 
+        # Basic text statistics
         words = user_input.split()
-        features[0] = len(words)
-        features[1] = user_input.count("?") / len(words)
-        features[2] = user_input.count("!") / len(words)
-        features[3] = len(user_input) / 100
-        features[4] = sum(
-            1
-            for word in words
-            if word.lower() in ["please", "thank", "thanks", "appreciate"]
-        ) / len(words)
-
-        features = features.astype(float)
+        features[0] = len(words)  # Word count
+        features[1] = sum(len(word) for word in words) / max(len(words), 1)  # Average word length
+        features[2] = user_input.count("?") / max(len(words), 1)  # Question mark frequency
+        features[3] = user_input.count("!") / max(len(words), 1)  # Exclamation mark frequency
+
+        # Sentiment analysis
+        blob = TextBlob(user_input)
+        features[4] = blob.sentiment.polarity  # Sentiment polarity (-1 to 1)
+        features[5] = blob.sentiment.subjectivity  # Subjectivity (0 to 1)
+
+        # Keyword detection
+        keywords = ["help", "explain", "understand", "confused", "clarify"]
+        features[6] = sum(word.lower() in keywords for word in words) / max(len(words), 1)
+
+        # Complexity indicators
+        features[7] = len(set(words)) / max(len(words), 1)  # Lexical diversity
+        features[8] = sum(len(word) > 6 for word in words) / max(len(words), 1)  # Proportion of long words
+
+        # Politeness indicator
+        polite_words = ["please", "thank", "thanks", "appreciate", "kindly"]
+        features[9] = sum(word.lower() in polite_words for word in words) / max(len(words), 1)
+
+        # spaCy processing
+        doc = self.nlp(user_input)
+
+        # Named Entity Recognition
+        features[10] = len(doc.ents) / max(len(words), 1)  # Named entity density
+
+        # Part-of-speech tagging
+        pos_counts = {pos: 0 for pos in ['NOUN', 'VERB', 'ADJ', 'ADV']}
+        for token in doc:
+            if token.pos_ in pos_counts:
+                pos_counts[token.pos_] += 1
+        features[11] = pos_counts['NOUN'] / max(len(words), 1)  # Noun density
+        features[12] = pos_counts['VERB'] / max(len(words), 1)  # Verb density
+
+        # Dependency parsing
+        features[13] = len([token for token in doc if token.dep_ == 'ROOT']) / max(len(words), 1)  # Main clause density
+
+        # Sentence complexity (using dependency parse tree depth)
+        def tree_depth(token):
+            return 1 + max((tree_depth(child) for child in token.children), default=0)
+
+        features[14] = sum(tree_depth(sent.root) for sent in doc.sents) / max(len(list(doc.sents)), 1)  # Average parse tree depth
+
+        # Additional features to match the environment's state dimension
+        features[15] = len([token for token in doc if token.is_stop]) / max(len(words), 1)  # Stop word density
+        features[16] = len([token for token in doc if token.is_punct]) / max(len(words), 1)  # Punctuation density
 
         return features

src/active_inference_forager/agents/philosophy_tutor_agent.py

@@ -2,13 +2,15 @@
 from typing import Dict, List
 from pydantic import Field
 from active_inference_forager.agents.generic_agent import GenericAgent
+import spacy
 
 
 class PhilosophyTutorAgent(GenericAgent):
     knowledge_base: Dict[str, Dict] = Field(default_factory=dict)
+    nlp: spacy.language.Language = Field(default_factory=lambda: spacy.load("en_core_web_sm"))
 
-    def __init__(self, state_dim: int, action_dim: int, **kwargs):
-        super().__init__(state_dim=state_dim, action_dim=action_dim, **kwargs)
+    def __init__(self, action_dim: int, **kwargs):
+        super().__init__(action_dim=action_dim, **kwargs)
         self.action_space = [
             "explain_concept",
             "ask_question",
@@ -67,25 +69,50 @@ def generate_response(self, action: str, state: np.ndarray) -> str:
             return "I'm not sure how to respond to that."
 
     def explain_philosophical_concept(self, state: np.ndarray) -> str:
-        # TODO: Implement logic to choose a concept based on the state
-        concept = "epistemology"  # Placeholder
+        # Use the state vector to choose a concept
+        concepts = list(self.knowledge_base['concepts'].keys())
+        concept_index = int(state[0] * len(concepts)) % len(concepts)
+        concept = concepts[concept_index]
         return f"Let me explain {concept}: {self.knowledge_base['concepts'][concept]}"
 
     def ask_socratic_question(self, state: np.ndarray) -> str:
-        # TODO: Implement logic to generate a relevant question based on the state
-        return "What do you think it means to truly know something?"
+        # Use the state vector to generate a relevant question
+        questions = [
+            "What do you think it means to truly know something?",
+            "How can we determine what is morally right or wrong?",
+            "What is the nature of reality, in your opinion?",
+            "How do you think we can achieve a just society?",
+        ]
+        question_index = int(state[1] * len(questions)) % len(questions)
+        return questions[question_index]
 
     def introduce_related_idea(self, state: np.ndarray) -> str:
-        # TODO: Implement logic to choose a related idea based on the state
-        return "Have you considered how this relates to the concept of free will?"
+        # Use the state vector to choose a related idea
+        ideas = [
+            "free will",
+            "consciousness",
+            "personal identity",
+            "the meaning of life",
+        ]
+        idea_index = int(state[2] * len(ideas)) % len(ideas)
+        return f"Have you considered how this relates to the concept of {ideas[idea_index]}?"
 
     def provide_example(self, state: np.ndarray) -> str:
-        # TODO: Implement logic to choose a relevant example based on the state
-        return "For instance, consider how we use logic in everyday decision-making..."
+        # Use the state vector to choose a relevant example
+        examples = [
+            "Consider how we use logic in everyday decision-making...",
+            "Think about how ethical considerations shape our laws and social norms...",
+            "Reflect on how our understanding of reality influences our actions...",
+            "Examine how our beliefs about knowledge affect our learning processes...",
+        ]
+        example_index = int(state[3] * len(examples)) % len(examples)
+        return examples[example_index]
 
     def suggest_thought_experiment(self, state: np.ndarray) -> str:
-        # TODO: Implement logic to choose a thought experiment based on the state
-        experiment = "The Cave"  # Placeholder
+        # Use the state vector to choose a thought experiment
+        experiments = list(self.knowledge_base['thought_experiments'].keys())
+        experiment_index = int(state[4] * len(experiments)) % len(experiments)
+        experiment = experiments[experiment_index]
         return f"Let's explore {experiment}: {self.knowledge_base['thought_experiments'][experiment]}"
 
     def acknowledge_limitation(self, state: np.ndarray) -> str:
@@ -95,9 +122,8 @@ def update_belief(self, state: np.ndarray):
         # We're now working directly with the state vector
         super().update_belief(state)
 
-    def process_user_input(self, state: np.ndarray) -> np.ndarray:
-        # This method now processes the state vector instead of a string
-        # We'll return the state as is, since it's already a numpy array
-        return state
+    def process_user_input(self, user_input: str) -> np.ndarray:
+        # Use the GenericAgent's process_user_input method
+        return super().process_user_input(user_input)
 
     # ... rest of the class implementation remains the same ...