smbot/ars.lua

-- Augmented Random Search
-- https://arxiv.org/abs/1803.07055
-- with some tweaks (lipschitz stuff) by myself.
-- i also added an option for graycode sampling,
-- borrowed from a (1+1) optimizer,
-- but i haven't yet found a case where it performs better.

local abs = math.abs
local exp = math.exp
local floor = math.floor
local insert = table.insert
local ipairs = ipairs
local max = math.max
local print = print

local Base = require "Base"

local nn = require "nn"
local normal = nn.normal
local prod = nn.prod
local uniform = nn.uniform
local zeros = nn.zeros

local util = require "util"
local argsort = util.argsort
local calc_mean_dev = util.calc_mean_dev
local normalize_sums = util.normalize_sums
local sign = util.sign

local Ars = Base:extend()

local exp_lut = {}
exp_lut[-1] = exp(-1)
exp_lut[0] = exp(0)
exp_lut[1] = exp(1)

local function collect_best_indices(scored, top, antithetic)
    -- select one (the best) reward of each pos/neg pair.
    local best_rewards
    if antithetic then
        best_rewards = {}
        for i = 1, #scored / 2 do
            local pos = scored[i * 2 - 1]
            local neg = scored[i * 2 - 0]
            best_rewards[i] = max(pos, neg)
        end
    else
        best_rewards = scored
    end

    local indices = argsort(best_rewards, function(a, b) return a > b end)

    for i = top + 1, #best_rewards do indices[i] = nil end
    return indices
end

local function kinda_lipschitz(dir, pos, neg, mid)
    -- based on the local lipschitz constant of a quadratic curve
    -- drawn through the 3 sampled points: positive, negative, and unperturbed.
    -- it kinda helps? there's probably a better function to base it around.
    local _, dev = calc_mean_dev(dir)
    local c0 = neg - mid
    local c1 = pos - mid
    local l0 = abs(3 * c1 + c0)
    local l1 = abs(c1 + 3 * c0)
    return max(l0, l1) / (2 * dev)
end

function Ars:init(dims, popsize, poptop, base_rate, sigma, antithetic,
                  momentum)
    self.dims = dims
    self.popsize = popsize or 4 + (3 * floor(log(dims)))
    base_rate = base_rate or 3/5 * (3 + log(dims)) / (dims * sqrt(dims))
    self.param_rate = base_rate
    self.sigma_rate = base_rate
    self.covar_rate = base_rate
    self.sigma = sigma or 1
    self.antithetic = antithetic == nil and true or antithetic
    self.momentum = momentum or 0

    self.poptop = poptop or popsize
    assert(self.poptop <= popsize)
    if self.antithetic then self.popsize = self.popsize * 2 end

    self._params = zeros(self.dims)
    if self.momentum > 0 then self.accum = zeros(self.dims) end

    self.evals = 0
end

function Ars:params(new_params)
    if new_params ~= nil then
        assert(#self._params == #new_params, "new parameters have the wrong size")
        for i, v in ipairs(new_params) do self._params[i] = v end
    end
    return self._params
end

function Ars:decay(param_decay, sigma_decay)
    if param_decay > 0 then
        for i, v in ipairs(self._params) do
            self._params[i] = v * (1 - self.param_rate * param_decay * self.sigma)
        end
    end
end

function Ars:ask(graycode)
    local asked = {}
    local noise = {}

    for i = 1, self.popsize do
        local asking = zeros(self.dims)
        local noisy = zeros(self.dims)
        asked[i] = asking
        noise[i] = noisy

        if self.antithetic and i % 2 == 0 then
            local old_noisy = noise[i - 1]
            for j, v in ipairs(old_noisy) do
                noisy[j] = -v
            end
        else
            if graycode ~= nil then
                for j = 1, self.dims do
                    noisy[j] = exp(-precision * uniform())
                end
                for j = 1, self.dims do
                    noisy[j] = uniform() < 0.5 and noisy[j] or -noisy[j]
                end
            else
                for j = 1, self.dims do
                    noisy[j] = self.sigma * normal()
                end
            end
        end

        for j, v in ipairs(self._params) do
            asking[j] = v + noisy[j]
        end
    end

    self.noise = noise
    return asked, noise
end

function Ars:tell(scored, unperturbed_score)
    local use_lips = unperturbed_score ~= nil and self.antithetic
    self.evals = self.evals + #scored
    if use_lips then self.evals = self.evals + 1 end

    local indices = collect_best_indices(scored, self.poptop, self.antithetic)

    local top_rewards = {}
    if self.antithetic then
        for _, ind in ipairs(indices) do
            insert(top_rewards, scored[ind * 2 - 1])
            insert(top_rewards, scored[ind * 2 - 0])
        end
    else
        -- ARS is built around antithetic sampling,
        -- but we can still do something without.
        -- this is getting to be very similar to SNES however.
        for _, ind in ipairs(indices) do insert(top_rewards, scored[ind]) end
        -- note: although this normalizes the scale, it's later
        --       re-normalized differently by reward_dev anyway.
        top_rewards = normalize_sums(top_rewards)
    end

    local step = zeros(self.dims)
    local _, reward_dev = calc_mean_dev(top_rewards)
    if reward_dev == 0 then reward_dev = 1 end

    if self.antithetic then
        for i, ind in ipairs(indices) do
            local pos = top_rewards[i * 2 - 1]
            local neg = top_rewards[i * 2 - 0]
            local reward = pos - neg
            if reward ~= 0 then
                local noisy = self.noise[ind * 2 - 1]
                if use_lips then
                    local lips = kinda_lipschitz(noisy, pos, neg, unperturbed_score)
                    reward = reward / lips / self.sigma
                else
                    reward = reward / reward_dev
                end

                for j, v in ipairs(noisy) do
                    step[j] = step[j] + reward * v / self.poptop
                end
            end
        end
    else
        for i, ind in ipairs(indices) do
            local reward = top_rewards[i] / reward_dev
            if reward ~= 0 then
                local noisy = self.noise[ind]

                for j, v in ipairs(noisy) do
                    step[j] = step[j] + reward * v / self.poptop
                end
            end
        end
    end

    if self.momentum > 0 then
        for i, v in ipairs(step) do
            self.accum[i] = self.momentum * self.accum[i] + v
            step[i] = v * exp_lut[sign(v) * sign(self.accum[i])]
        end
    end

    for i, v in ipairs(self._params) do
        self._params[i] = v + self.param_rate * step[i]
    end

    self.noise = nil

    return step
end

return {
    collect_best_indices = collect_best_indices,
    Ars = Ars,
}