foxhunt/ml/tests/ppo_continuous_policy_unit_test.rs

//! Unit tests for PPO Continuous Policy
//!
//! Tests continuous action spaces, policy network, and action sampling.

use anyhow::Result;
use candle_core::{Device, Tensor};
use ml::ppo::continuous_policy::{ContinuousPolicyConfig, ContinuousPolicyNetwork};

#[test]
fn test_continuous_policy_creation() -> Result<()> {
    let device = Device::Cpu;
    let config = ContinuousPolicyConfig {
        input_dim: 64,
        hidden_dim: 128,
        action_dim: 4,
        learning_rate: 3e-4,
    };

    let policy = ContinuousPolicyNetwork::new(config, &device)?;

    // Verify policy was created
    assert!(policy.input_dim() == 64);
    assert!(policy.action_dim() == 4);

    Ok(())
}

#[test]
fn test_continuous_policy_forward() -> Result<()> {
    let device = Device::Cpu;
    let batch_size = 8;
    let input_dim = 64;
    let action_dim = 4;

    let config = ContinuousPolicyConfig {
        input_dim,
        hidden_dim: 128,
        action_dim,
        learning_rate: 3e-4,
    };

    let policy = ContinuousPolicyNetwork::new(config, &device)?;

    // Create state input
    let state = Tensor::randn(0.0f32, 1.0, (batch_size, input_dim), &device)?;

    // Forward pass returns (mean, std)
    let (mean, std) = policy.forward(&state)?;

    // Verify shapes
    assert_eq!(mean.dims(), &[batch_size, action_dim]);
    assert_eq!(std.dims(), &[batch_size, action_dim]);

    // Verify std is positive
    let std_min = std.min(0)?.min(0)?.to_scalar::<f32>()?;
    assert!(std_min > 0.0, "Standard deviation should be positive");

    Ok(())
}

#[test]
fn test_continuous_policy_action_sampling() -> Result<()> {
    let device = Device::Cpu;
    let batch_size = 4;
    let input_dim = 32;
    let action_dim = 2;

    let config = ContinuousPolicyConfig {
        input_dim,
        hidden_dim: 64,
        action_dim,
        learning_rate: 3e-4,
    };

    let policy = ContinuousPolicyNetwork::new(config, &device)?;
    let state = Tensor::randn(0.0f32, 1.0, (batch_size, input_dim), &device)?;

    // Sample actions
    let action = policy.sample_action(&state)?;

    // Verify action shape
    assert_eq!(action.dims(), &[batch_size, action_dim]);

    // Actions should be finite
    let action_max = action.abs()?.max(0)?.max(0)?.to_scalar::<f32>()?;
    assert!(action_max.is_finite());

    Ok(())
}

#[test]
fn test_continuous_policy_deterministic_mode() -> Result<()> {
    let device = Device::Cpu;
    let batch_size = 2;
    let input_dim = 32;
    let action_dim = 2;

    let config = ContinuousPolicyConfig {
        input_dim,
        hidden_dim: 64,
        action_dim,
        learning_rate: 3e-4,
    };

    let policy = ContinuousPolicyNetwork::new(config, &device)?;
    let state = Tensor::ones((batch_size, input_dim), candle_core::DType::F32, &device)?;

    // In deterministic mode, should return mean
    let (mean, _) = policy.forward(&state)?;
    let action = policy.deterministic_action(&state)?;

    // Action should equal mean in deterministic mode
    let diff = (&action - &mean)?.abs()?.sum_all()?.to_scalar::<f32>()?;
    assert!(diff < 1e-5, "Deterministic action should equal mean");

    Ok(())
}

#[test]
fn test_continuous_policy_log_prob() -> Result<()> {
    let device = Device::Cpu;
    let batch_size = 4;
    let input_dim = 32;
    let action_dim = 2;

    let config = ContinuousPolicyConfig {
        input_dim,
        hidden_dim: 64,
        action_dim,
        learning_rate: 3e-4,
    };

    let policy = ContinuousPolicyNetwork::new(config, &device)?;
    let state = Tensor::randn(0.0f32, 1.0, (batch_size, input_dim), &device)?;
    let action = Tensor::randn(0.0f32, 1.0, (batch_size, action_dim), &device)?;

    // Compute log probability
    let log_prob = policy.log_prob(&state, &action)?;

    // Verify shape
    assert_eq!(log_prob.dims(), &[batch_size]);

    // Log probabilities should be negative or zero
    let log_prob_max = log_prob.max(0)?.to_scalar::<f32>()?;
    assert!(log_prob_max <= 0.01, "Log probabilities should be ≤ 0");

    Ok(())
}

#[test]
fn test_continuous_policy_entropy() -> Result<()> {
    let device = Device::Cpu;
    let batch_size = 4;
    let input_dim = 32;
    let action_dim = 2;

    let config = ContinuousPolicyConfig {
        input_dim,
        hidden_dim: 64,
        action_dim,
        learning_rate: 3e-4,
    };

    let policy = ContinuousPolicyNetwork::new(config, &device)?;
    let state = Tensor::randn(0.0f32, 1.0, (batch_size, input_dim), &device)?;

    // Compute entropy
    let entropy = policy.entropy(&state)?;

    // Entropy should be positive (Gaussian entropy > 0)
    let entropy_min = entropy.min(0)?.to_scalar::<f32>()?;
    assert!(entropy_min > 0.0, "Entropy should be positive");

    Ok(())
}

#[test]
fn test_continuous_policy_gradient_flow() -> Result<()> {
    let device = Device::Cpu;
    let batch_size = 4;
    let input_dim = 32;
    let action_dim = 2;

    let config = ContinuousPolicyConfig {
        input_dim,
        hidden_dim: 64,
        action_dim,
        learning_rate: 3e-4,
    };

    let policy = ContinuousPolicyNetwork::new(config, &device)?;
    let state = Tensor::randn(0.0f32, 1.0, (batch_size, input_dim), &device)?;
    let action = policy.sample_action(&state)?;

    // Compute log prob and loss
    let log_prob = policy.log_prob(&state, &action)?;
    let loss = log_prob.sum_all()?;

    // Verify backward pass works
    loss.backward()?;

    Ok(())
}

#[test]
fn test_continuous_policy_action_bounds() -> Result<()> {
    let device = Device::Cpu;
    let batch_size = 10;
    let input_dim = 32;
    let action_dim = 2;

    let config = ContinuousPolicyConfig {
        input_dim,
        hidden_dim: 64,
        action_dim,
        learning_rate: 3e-4,
    };

    let policy = ContinuousPolicyNetwork::new(config, &device)?;
    let state = Tensor::randn(0.0f32, 1.0, (batch_size, input_dim), &device)?;

    // Sample many actions
    for _ in 0..100 {
        let action = policy.sample_action(&state)?;

        // Actions should be within reasonable bounds (e.g., ±10)
        let action_max = action.abs()?.max(0)?.max(0)?.to_scalar::<f32>()?;
        assert!(action_max < 100.0, "Actions should not be extreme");
    }

    Ok(())
}

#[test]
fn test_continuous_policy_different_action_dims() -> Result<()> {
    let device = Device::Cpu;
    let batch_size = 4;
    let input_dim = 64;

    // Test various action dimensions
    for action_dim in [1, 2, 4, 8] {
        let config = ContinuousPolicyConfig {
            input_dim,
            hidden_dim: 128,
            action_dim,
            learning_rate: 3e-4,
        };

        let policy = ContinuousPolicyNetwork::new(config, &device)?;
        let state = Tensor::randn(0.0f32, 1.0, (batch_size, input_dim), &device)?;

        let (mean, std) = policy.forward(&state)?;

        assert_eq!(mean.dim(1)?, action_dim);
        assert_eq!(std.dim(1)?, action_dim);
    }

    Ok(())
}

#[test]
fn test_continuous_policy_consistency() -> Result<()> {
    let device = Device::Cpu;
    let batch_size = 2;
    let input_dim = 32;
    let action_dim = 2;

    let config = ContinuousPolicyConfig {
        input_dim,
        hidden_dim: 64,
        action_dim,
        learning_rate: 3e-4,
    };

    let policy = ContinuousPolicyNetwork::new(config, &device)?;
    let state = Tensor::ones((batch_size, input_dim), candle_core::DType::F32, &device)?;

    // Same input should give same mean (deterministic forward)
    let (mean1, _) = policy.forward(&state)?;
    let (mean2, _) = policy.forward(&state)?;

    let diff = (&mean1 - &mean2)?.abs()?.sum_all()?.to_scalar::<f32>()?;
    assert!(diff < 1e-6, "Forward pass should be deterministic");

    Ok(())
}