Add symbolic regression experiment.

offchain/src/bin/gp_spec_tester.rs

@@ -0,0 +1,36 @@
+use anyhow::Result;
+use rand::thread_rng;
+use rand::Rng;
+
+// Suppose you have these in your library:
+use offchain::gp::generate_spec::ranmdom_code_fixed;
+use offchain::gp::mutation::mutate_by_index;
+use offchain::compiler::ast::UntypedAst;
+
+fn main() -> Result<()> {
+    let mut rng = thread_rng();
+
+    // We'll produce, say, 3 random ASTs ("subjects").
+    let num_subjects = 3;
+    let max_points = 6;
+
+    for i in 0..num_subjects {
+        // Generate a random AST
+        let original = ranmdom_code_fixed(&mut rng, max_points);
+        println!("=== Subject #{} ===", i);
+        println!("Original AST:\n{:#?}", original);
+
+        // Let's do 2 successive mutations on it
+        let mut prev = original.clone();
+        for mut_idx in 0..2 {
+            let mutated = mutate_by_index(&prev, &mut rng, max_points);
+            println!("Mutation #{} result:\n{:#?}", mut_idx, mutated);
+            prev = mutated; // so the next iteration mutates from the last result
+        }
+
+        println!("=== End of Subject #{} ===\n", i);
+    }
+
+    println!("Done generating & mutating random ASTs!");
+    Ok(())
+}

offchain/src/bin/symreg_experiment.rs

@@ -0,0 +1,170 @@
+// src/bin/symreg_experiment.rs
+
+use anyhow::Result;
+use rand::{thread_rng, Rng};
+use offchain::compiler::ast::{UntypedAst, Push3Ast};
+use offchain::helpers::artifact::get_creation_code;
+use offchain::runner::revm_runner::{
+    EvmRunner, 
+    Push3InterpreterInputs, 
+    Push3InterpreterOutputs
+};
+use offchain::gp::generate_spec::{ranmdom_code_fixed};
+use offchain::gp::mutation::mutate_by_index;
+use offchain::compiler::push3_describtor::make_sublist_descriptor;
+use ethers::types::U256;
+use ethers::abi::{encode, Token};
+
+// 1) Generate (x, y) samples for f(x) = 3x^2 + x + 3
+fn generate_samples() -> Vec<(i32, i32)> {
+    let mut samples = Vec::new();
+    for x in -5..=5 {
+        let y = 3 * x * x + x + 3;
+        samples.push((x, y));
+    }
+    samples
+}
+
+// 2) Evaluate a single AST on a single input `x`, returning the integer
+//    that ends up on top of the final int stack (or 0 if empty).
+fn evaluate_ast_on_x(
+    runner: &mut EvmRunner,
+    ast: &UntypedAst,
+    x: i32,
+) -> i32 {
+    // a) Convert AST => push3 bytecode
+    let code_bytes = ast.to_bytecode();
+    let code_len = code_bytes.len() as u32;
+
+    // b) Build a sublist descriptor
+    let descriptor = make_sublist_descriptor(0, code_len);
+
+    // c) Provide x in `init_int_stack`
+    let inputs = Push3InterpreterInputs {
+        code: code_bytes,
+        init_code_stack: Vec::new(),
+        init_exec_stack: vec![descriptor],
+        init_int_stack: vec![x as i128],
+    };
+
+    // d) Actually run the interpreter
+    let outputs = match runner.run_interpreter(&inputs) {
+        Ok(o) => o,
+        Err(_) => {
+            // If the call fails or reverts, return 0
+            return 0;
+        }
+    };
+
+    // e) If final_int_stack is empty => 0, else top item
+    if outputs.final_int_stack.is_empty() {
+        0
+    } else {
+        *outputs.final_int_stack.last().unwrap() as i32
+    }
+}
+
+// 3) Evaluate an AST on all (x, y) samples => compute MSE
+fn evaluate_fitness(
+    runner: &mut EvmRunner,
+    ast: &UntypedAst,
+    samples: &[(i32, i32)]
+) -> f64 {
+    let mut error_sum = 0.0;
+    for &(x, target_y) in samples {
+        let predicted = evaluate_ast_on_x(runner, ast, x);
+        let diff = (predicted - target_y) as f64;
+        error_sum += diff * diff;
+    }
+    error_sum / samples.len() as f64
+}
+
+fn main() -> Result<()> {
+    // A) Generate samples for f(x)=3x^2 + x + 3
+    let samples = generate_samples();
+
+    // B) Create an ephemeral EVM runner (Push3Interpreter)
+    let creation_hex_filename = "../onchain/out/Push3Interpreter.sol/Push3Interpreter.json";
+    let creation_bytes = get_creation_code(creation_hex_filename)?;
+    let mut runner = EvmRunner::new(creation_bytes)?;
+
+    // C) Setup GP parameters
+    let pop_size = 1000;       // population size
+    let max_points = 15;      // max "size" for random code
+    let generations = 100;    // how many generations we run
+    let early_stop_threshold = 1.0; // if error < this => we can stop
+
+    let mut rng = thread_rng();
+
+    // D) Initialize random population
+    let mut population: Vec<UntypedAst> = (0..pop_size)
+        .map(|_| ranmdom_code_fixed(&mut rng, max_points))
+        .collect();
+
+    // E) Main GP Loop
+    for gen in 0..generations {
+        // 1) Evaluate each => store (ast, error)
+        let mut scored: Vec<(UntypedAst, f64)> = population.into_iter()
+            .map(|ast| {
+                let err = evaluate_fitness(&mut runner, &ast, &samples);
+                (ast, err)
+            })
+            .collect();
+
+        // 2) Sort ascending by error
+        scored.sort_by(|a, b| a.1.partial_cmp(&b.1).unwrap());
+
+        // 3) Print best error
+        let best_err = scored[0].1;
+        println!("Generation {gen}, best error = {best_err}");
+
+        // 4) Check if best is below threshold => early stop
+        if best_err < early_stop_threshold {
+            println!("Best error < {early_stop_threshold}, stopping early!");
+            population = scored.into_iter().map(|(a, _)| a).collect();
+            break;
+        }
+
+        // 5) Keep best half
+        let half = scored.len() / 2;
+        let survivors = &scored[..half];
+
+        // 6) Refill population: half mutated, half fresh random
+        let mut new_pop = Vec::with_capacity(scored.len());
+        // a) keep survivors
+        for &(ref ast, _) in survivors {
+            new_pop.push(ast.clone());
+        }
+
+        // b) how many to add?
+        let remainder_size = pop_size - new_pop.len();
+        // we do half mutated, half new random
+        let mutated_count = remainder_size / 2;
+        let random_count = remainder_size - mutated_count;
+
+        // c) add mutated
+        for _ in 0..mutated_count {
+            let idx = rng.gen_range(0..half);
+            let mutated = mutate_by_index(&survivors[idx].0, &mut rng, max_points);
+            new_pop.push(mutated);
+        }
+
+        // d) add brand-new random
+        for _ in 0..random_count {
+            let fresh = ranmdom_code_fixed(&mut rng, max_points);
+            new_pop.push(fresh);
+        }
+
+        // store for next generation
+        population = new_pop;
+    }
+
+    // F) Final Print
+    println!("\n=== Final Population ===");
+    for (i, ast) in population.iter().enumerate() {
+        let err = evaluate_fitness(&mut runner, ast, &samples);
+        println!("Subject #{i}, err={err}, ast={:?}", ast);
+    }
+
+    Ok(())
+}

offchain/src/gp/generate_spec.rs

@@ -0,0 +1,121 @@
+use crate::compiler::ast::{UntypedAst, OpCode};
+use rand::Rng;
+use rand::prelude::SliceRandom;
+
+#[derive(Clone, Debug)]
+pub enum InstructionAtom {
+    /// A normal opcode, e.g. `Plus`, `Minus`, `Dup`, etc.
+    Opcode(OpCode),
+    /// An ephemeral random constant placeholder. 
+    /// If chosen, we generate a random literal on the spot.
+    EphemeralInt,
+    // If you want ephemeral floats, booleans, etc., add more variants
+}
+
+/// A small struct to hold our entire “instruction set.”
+#[derive(Clone, Debug)]
+pub struct InstructionSet {
+    pub atoms: Vec<InstructionAtom>,
+}
+
+impl InstructionSet {
+    pub fn new_default() -> Self {
+        // Example: Some opcodes + ephemeral constant
+        // Tweak as you like
+        use InstructionAtom::*;
+        use OpCode::*;
+
+        Self {
+            atoms: vec![
+                Opcode(Noop),
+                Opcode(Plus),
+                Opcode(Minus),
+                Opcode(Mult),
+                Opcode(Dup),
+                Opcode(Pop),
+                EphemeralInt,
+            ],
+        }
+    }
+    
+    /// Pick a random atom from this set.
+    /// If it's `EphemeralInt`, we produce `UntypedAst::IntLiteral(...)`.
+    /// If it's `Opcode(...)`, we produce `UntypedAst::Instruction(...)`.
+    pub fn random_atom_as_ast(&self, rng: &mut impl Rng) -> UntypedAst {
+        let idx = rng.gen_range(0..self.atoms.len());
+        match &self.atoms[idx] {
+            InstructionAtom::Opcode(op) => UntypedAst::Instruction(op.clone()),
+            InstructionAtom::EphemeralInt => {
+                // For ephemeral int, produce a random literal in some range
+                let val = rng.gen_range(-10..10);
+                UntypedAst::IntLiteral(val)
+            }
+        }
+    }
+}
+
+pub fn ranmdom_code_fixed(rng: &mut impl Rng, max_points: usize) -> UntypedAst {
+    let instr_set = InstructionSet::new_default();
+    return random_code(rng, &instr_set, max_points);
+}
+
+/// The "entry point": random_code(max_points).
+/// 1) Choose actual_points in [1..=max_points].
+/// 2) Return the result of random_code_with_size.
+pub fn random_code(rng: &mut impl Rng, instr_set: &InstructionSet, max_points: usize) -> UntypedAst {
+    let actual_points = rng.gen_range(1..=max_points);
+    random_code_with_size(rng, instr_set, actual_points)
+}
+
+/// The main logic: 
+///   If points == 1 => pick a single instruction/ephemeral 
+///   Else => 
+///       - We "decompose" (points - 1) among sub-codes
+///       - For each sub-points, we recursively call random_code_with_size
+///       - We return a Sublist of those children, possibly in random order
+pub fn random_code_with_size(
+    rng: &mut impl Rng,
+    instr_set: &InstructionSet,
+    points: usize,
+) -> UntypedAst {
+    use UntypedAst::*;
+
+    // Base case
+    if points == 1 {
+        // Choose a random "atom" from instr_set
+        return instr_set.random_atom_as_ast(rng);
+    } 
+
+    // If `points > 1`, let's produce a Sublist
+    // We'll break (points - 1) into sub-points via `decompose`
+    let subpoints_list = decompose(rng, points - 1, points - 1);
+    // subpoints_list is e.g. [2, 3, 5] and sums to (points-1)
+
+    // Then for each "subpoints", we do random_code_with_size 
+    let mut sub_asts: Vec<UntypedAst> = subpoints_list
+        .into_iter()
+        .map(|sp| random_code_with_size(rng, instr_set, sp))
+        .collect();
+
+    // The spec says "Return a list containing the results, in random order"
+    // so let's shuffle sub_asts
+    sub_asts.shuffle(rng);
+
+    Sublist(sub_asts)
+}
+
+/// Decompose a number into random parts. 
+/// 
+///  - If number is 1 or max_parts is 1 => return [number].
+///  - Otherwise pick a random split "this_part" and recurse on the remainder.
+fn decompose(rng: &mut impl Rng, number: usize, max_parts: usize) -> Vec<usize> {
+    if number == 1 || max_parts == 1 {
+        return vec![number];
+    }
+    // pick a random "this_part" in [1 .. (number-1)]
+    let this_part = rng.gen_range(1..number);
+    let mut remainder = decompose(rng, number - this_part, max_parts - 1);
+    let mut result = vec![this_part];
+    result.append(&mut remainder);
+    result
+}

offchain/src/gp/mod.rs

@@ -1,3 +1,5 @@
 pub mod population;
 pub mod genetic_ops;
 pub mod generate;
+pub mod generate_spec;
+pub mod mutation;

offchain/src/gp/mutation.rs

@@ -0,0 +1,86 @@
+use rand::Rng;
+use crate::compiler::ast::UntypedAst;
+use crate::gp::generate_spec::ranmdom_code_fixed; // or push3-based generator
+
+/// A "path" is a list of indices leading from the root to a child.
+/// For example, [] is the root, [0] is the root's first child, [0,1] is that child's second child, etc.
+pub type Path = Vec<usize>;
+
+/// Enumerate all nodes in a DFS order (root first).
+/// For each node, we store the `Path` (list of child indices).
+pub fn enum_nodes_dfs(ast: &UntypedAst) -> Vec<Path> {
+    let mut paths = Vec::new();
+    dfs_helper(ast, &mut paths, &mut vec![]);
+    paths
+}
+
+/// Recursive helper: `current_path` is a stack we push child indices onto.
+/// `paths` is the global list of final results.
+fn dfs_helper(ast: &UntypedAst, paths: &mut Vec<Path>, current_path: &mut Vec<usize>) {
+    // Record the current path as a valid node
+    paths.push(current_path.clone());
+
+    match ast {
+        UntypedAst::Sublist(children) => {
+            // For each child, push its index, recurse, then pop
+            for (i, child) in children.iter().enumerate() {
+                current_path.push(i);
+                dfs_helper(child, paths, current_path);
+                current_path.pop();
+            }
+        }
+        UntypedAst::IntLiteral(_) | UntypedAst::Instruction(_) => {
+            // Leaf node => no children
+        }
+    }
+}
+
+/// Replace the node at `path` in `original` with `replacement`, returning a new AST.
+/// If `path` is empty => we replace the root entirely.
+pub fn replace_subtree(
+    original: &UntypedAst,
+    path: &[usize],
+    replacement: UntypedAst
+) -> UntypedAst {
+    if path.is_empty() {
+        // means we’re replacing the root
+        return replacement;
+    }
+    // otherwise, we need to walk the sublist chain
+    match original {
+        UntypedAst::Sublist(children) => {
+            // let path_head = path[0], path_tail = path[1..]
+            let (first_idx, tail_path) = (path[0], &path[1..]);
+            let mut new_children = children.clone();
+            new_children[first_idx] = replace_subtree(&children[first_idx], tail_path, replacement);
+            UntypedAst::Sublist(new_children)
+        }
+        // If the path is not empty but we’re not in a Sublist => error or do nothing
+        UntypedAst::IntLiteral(_) | UntypedAst::Instruction(_) => {
+            // Possibly, if the path is invalid, we ignore. Or we treat it as "replace anyway."
+            // Here, we might treat it as an error or a no-op.
+            original.clone()
+        }
+    }
+}
+
+/// Produces a new AST by choosing exactly one node in `original` at random
+/// (by enumerating them), and replacing it with a new random subtree.
+pub fn mutate_by_index(
+    original: &UntypedAst,
+    rng: &mut impl Rng,
+    max_points: usize,
+) -> UntypedAst {
+    // 1) Enumerate all nodes => get a vector of `Path`
+    let all_paths = enum_nodes_dfs(original);
+    // pick a random path
+    let idx = rng.gen_range(0..all_paths.len());
+    let chosen_path = &all_paths[idx];
+
+    // 2) Generate a new subtree with our "fixed" random code generator
+    let new_subtree = ranmdom_code_fixed(rng, max_points);
+
+    // 3) Replace the subtree at `chosen_path` in `original` with `new_subtree`
+    replace_subtree(original, chosen_path, new_subtree)
+}
+