ReCo.jl/src/RL/LocalCOMEnv.jl

export LocalCOMEnv

using ..ReCo: Particle

struct LocalCOMEnv <: Env
    shared::EnvSharedProps

    distance_state_space::Vector{Interval}
    direction_angle_state_space::Vector{Interval}

    function LocalCOMEnv(;
        n_distance_states::Int64=3, n_direction_angle_states::Int64=3, args
    )
        @assert n_direction_angle_states > 1

        direction_angle_state_space = gen_angle_state_space(n_direction_angle_states)

        min_distance = 0.0
        max_distance = args.skin_r

        distance_state_space = gen_distance_state_space(
            min_distance, max_distance, n_distance_states
        )

        n_states = n_distance_states * n_direction_angle_states + 1

        state_space = Vector{SVector{2,Interval}}(undef, n_states - 1)

        ind = 1
        for distance_state in distance_state_space
            for direction_angle_state in direction_angle_state_space
                state_space[ind] = SVector(distance_state, direction_angle_state)
                ind += 1
            end
        end
        # Last state is when no particle is in the skin radius

        shared = EnvSharedProps(n_states, state_space)

        return new(shared, distance_state_space, direction_angle_state_space)
    end
end

mutable struct LocalCOMEnvHelper <: EnvHelper
    shared::EnvHelperSharedProps

    vec_to_neighbour_sums::Vector{SVector{2,Float64}}
    n_neighbours::Vector{Int64}

    distances_to_local_center_of_mass::Vector{Float64}
    max_distance_to_local_center_of_mass::Float64

    add_shape_reward_term::Bool

    center_of_mass::SVector{2,Float64}
    gyration_tensor_eigvec_to_smaller_eigval::SVector{2,Float64}
    gyration_tensor_eigvec_to_bigger_eigval::SVector{2,Float64}

    half_box_len::Float64
    max_elliptical_distance::Float64

    function LocalCOMEnvHelper(shared::EnvHelperSharedProps, half_box_len::Float64, skin_r)
        max_elliptical_distance =
            sqrt(2) * half_box_len / shared.goal_gyration_tensor_eigvals_ratio

        max_distance_to_local_center_of_mass = skin_r

        return new(
            shared,
            fill(SVector(0.0, 0.0), shared.n_particles),
            fill(0, shared.n_particles),
            zeros(shared.n_particles),
            max_distance_to_local_center_of_mass,
            false,
            SVector(0.0, 0.0),
            SVector(0.0, 0.0),
            SVector(0.0, 0.0),
            half_box_len,
            max_elliptical_distance,
        )
    end
end

function gen_env_helper(::LocalCOMEnv, env_helper_shared::EnvHelperSharedProps; args)
    return LocalCOMEnvHelper(env_helper_shared, args.half_box_len, args.skin_r)
end

function pre_integration_hook(env_helper::LocalCOMEnvHelper)
    @simd for id in 1:(env_helper.shared.n_particles)
        env_helper.vec_to_neighbour_sums[id] = SVector(0.0, 0.0)
        env_helper.n_neighbours[id] = 0
    end

    return nothing
end

function state_update_helper_hook(
    env_helper::LocalCOMEnvHelper, id1::Int64, id2::Int64, r⃗₁₂::SVector{2,Float64}
)
    env_helper.vec_to_neighbour_sums[id1] += r⃗₁₂
    env_helper.vec_to_neighbour_sums[id2] -= r⃗₁₂

    env_helper.n_neighbours[id1] += 1
    env_helper.n_neighbours[id2] += 1

    return nothing
end

function state_update_hook(env_helper::LocalCOMEnvHelper, particles::Vector{Particle})
    n_particles = env_helper.shared.n_particles

    @turbo for id in 1:n_particles
        env_helper.shared.old_states_ind[id] = env_helper.shared.states_ind[id]
    end

    env = env_helper.shared.env

    distance_to_local_center_of_mass_sum = 0.0

    for id in 1:n_particles
        n_neighbours = env_helper.n_neighbours[id]

        if n_neighbours == 0
            state_ind = env.shared.n_states

            distance_to_local_center_of_mass_sum +=
                env_helper.max_distance_to_local_center_of_mass
        else
            vec_to_local_center_of_mass =
                env_helper.vec_to_neighbour_sums[id] / n_neighbours

            distance = ReCo.norm2d(vec_to_local_center_of_mass)

            env_helper.distances_to_local_center_of_mass[id] = distance

            distance_to_local_center_of_mass_sum += distance

            distance_state = find_state_interval(distance, env.distance_state_space)

            si, co = sincos(particles[id].φ)

            direction_angle = ReCo.angle2(SVector(co, si), vec_to_local_center_of_mass)

            direction_angle_state = find_state_interval(
                direction_angle, env.direction_angle_state_space
            )

            state = SVector{2,Interval}(distance_state, direction_angle_state)
            state_ind = find_state_ind(state, env.shared.state_space)
        end

        env_helper.shared.states_ind[id] = state_ind
    end

    mean_distance_to_local_center_of_mass =
        distance_to_local_center_of_mass_sum / n_particles
    env_helper.add_shape_reward_term =
        mean_distance_to_local_center_of_mass /
        env_helper.max_distance_to_local_center_of_mass < 0.32

    if env_helper.add_shape_reward_term
        #println(mean_distance_to_local_center_of_mass / env_helper.max_distance_to_local_center_of_mass) # TODO: Remove
    end

    env_helper.center_of_mass = ReCo.center_of_mass(particles, env_helper.half_box_len)

    v1, v2 = ReCo.gyration_tensor_eigvecs(
        particles, env_helper.half_box_len, env_helper.center_of_mass
    )

    env_helper.gyration_tensor_eigvec_to_smaller_eigval = v1
    env_helper.gyration_tensor_eigvec_to_bigger_eigval = v2

    return nothing
end

"""
    minimizing_reward(value::Float64, max_value::Float64)

Returns the reward such that it is 0 for value=max_value and 1 for value=0.
"""
function minimizing_reward(value::Float64, max_value::Float64)
    return (max_value - value) / (max_value + value)
end

function update_reward!(env::LocalCOMEnv, env_helper::LocalCOMEnvHelper, particle::Particle)
    id = particle.id

    normalization = env_helper.shared.n_particles # TODO: Add factor from steps
    n_neighbours = env_helper.n_neighbours[id]

    if n_neighbours == 0
        env.shared.reward = 0.0
    else
        reward = minimizing_reward(
            env_helper.distances_to_local_center_of_mass[id],
            env_helper.max_distance_to_local_center_of_mass,
        )

        if env_helper.add_shape_reward_term
            elliptical_distance = ReCo.elliptical_distance(
                particle.c,
                env_helper.center_of_mass,
                env_helper.gyration_tensor_eigvec_to_smaller_eigval,
                env_helper.gyration_tensor_eigvec_to_bigger_eigval,
                env_helper.shared.goal_gyration_tensor_eigvals_ratio,
                env_helper.half_box_len,
            )

            reward += unnormalized_reward(
                elliptical_distance,
                env_helper.max_elliptical_distance, # TODO: Fix sq
            )

            # println(elliptical_distance / env_helper.max_elliptical_distance) # TODO: Remove
        end

        env.shared.reward = reward / normalization
    end

    return nothing
end