→ ↺ λ

Micmac Manual

Table of Contents

• 1 Introduction
  • 1.1 Overview
  • 1.2 Links
• 2 Graph Search
  • 2.1 UCT
• 3 Metropolis Hastings
• 4 Game Theory

[in package MICMAC]

• [system] "micmac"
  • Version: 0.0.2
  • Description: Micmac is mainly a library of graph search algorithms such as alpha-beta, UCT and beam search, but it also has some MCMC and other slightly unrelated stuff.
  • Licence: MIT, see COPYING.
  • Author: Gábor Melis mailto:mega@retes.hu
  • Mailto: mega@retes.hu
  • Homepage: http://melisgl.github.io/mgl-gpr
  • Bug tracker: https://github.com/melisgl/mgl-gpr/issues
  • Source control: GIT
  • Depends on: alexandria, mgl-pax

← ↑ → ↺ λ

1 Introduction

← ↑ → ↺ λ

1.1 Overview

MICMAC is a Common Lisp library by Gábor Melis focusing on graph search algorithms.

← ↑ → ↺ λ

1.2 Links

Here is the official repository and the HTML documentation for the latest version.

← ↑ → ↺ λ

2 Graph Search

• [function] alpha-beta state &key (depth 0) alpha beta call-with-action maybe-evaluate-state list-actions record-best

  Alpha-beta pruning for two player, zero-sum maximax (like minimax but both players maximize and the score is negated when passed between depths). Return the score of the game state from the point of view of the player to move at depth and as the second value the list of actions of the principal variant.

  call-with-action is a function of (state depth action FN). It carries out action (returned by list-actions or nil) to get the state corresponding to depth and calls FN with that state. It may destructively modify state provided it undoes the damage after FN returns. call-with-action is called with nil as action for the root of the tree, in this case state need not be changed. FN returns the same kinds of values as alpha-beta. They may be useful for logging.

  maybe-evaluate-state is a function of (state depth). If state at depth is a terminal node then it returns the score from the point of view of the player to move and as the second value a list of actions that lead from state to the position that was evaluated. The list of actions is typically empty. If we are not at a terminal node then maybe-evaluate-state returns nil.

  list-actions is a function of (state depth) and returns a non-empty list of legal candidate moves for non-terminal nodes. Actions are tried in the order list-actions returns them: stronger moves

  call-with-action, maybe-evaluate-state and list-actions are mandatory.

  record-best, if non-nil, is a function of (depth score actions). It is called when at depth a new best action is found. actions is a list of all the actions in the principle variant corresonding to the newly found best score. record-best is useful for graceful degradation in case of timeout.

  alpha and beta are typically nil (equivalent to -infinity, +infinity) but any real number is allowed if the range of scores can be boxed.

  See test/test-alpha-beta.lisp for an example.

• [function] beam-search start-nodes &key max-depth (n-solutions 1) (beam-width (length start-nodes)) expand-node-fn expand-beam-fn score-fn upper-bound-fn solutionp-fn (finishedp-fn solutionp-fn)

  In a graph, search for nodes that with the best scores with beam search. That is, starting from start-nodes perform a breadth-first search but at each depth only keep beam-width number of nodes with the best scores. Keep the best n-solutions (at most) complete solutions. Discard nodes known to be unable to get into the best n-solutions (due to upper-bound-fn). Finally, return the solutions and the active nodes (the beam) as adjustable arrays sorted by score in descending order.

  start-nodes (a sequence of elements of arbitrary type). score-fn, upper-bound-fn, solutionp-fn, finishedp-fn are all functions of one argument: the node. solutionp-fn checks whether a node represents a complete solution (i.e. some goal is reached). score-fn returns a real number that's to be maximized, it's only called for node for which solutionp-fn returned true. upper-bound-fn (if not nil) returns a real number that equal or greater than the score of all solutions reachable from that node. finishedp-fn returns true iff there is nowhere to go from the node.

  expand-node-fn is also a function of a single node argument. It returns a sequence of nodes to 'one step away' from its argument node. expand-beam-fn is similar, but it takes a vector of nodes and returns all nodes one step away from any of them. It's enough provide either expand-node-fn or expand-beam-fn. The purpose of expand-beam-fn. is to allow more efficient, batch-like operations.

  See test/test-beam-search.lisp for an example.

• [function] parallel-beam-search start-node-seqs &key max-depth (n-solutions 1) beam-width expand-node-fn expand-beams-fn score-fn upper-bound-fn solutionp-fn (finishedp-fn solutionp-fn)

  This is very much like beam-search except it solves a number of instances of the same search problem starting from different sets of nodes. The sole purpose of parallel-beam-search is to amortize the cost expand-beam-fn if possible.

  expand-beams-fn is called with sequence of beams (i.e. it's a sequence of sequence of nodes) and it must return another sequence of sequences of nodes. Each element of the returned sequence is the reachable nodes of the nodes in the corresponding element of its argument sequence.

  parallel-beam-search returns a sequence of solutions sequences, and a sequence of active node sequences.

  See test/test-beam-search.lisp for an example.

← ↑ → ↺ λ

2.1 UCT

[in package MICMAC.UCT]

uct Monte Carlo tree search. This is what makes current Go programs tick. And Hex programs as well, for that matter. This is a cleanup and generalization of code originally created in course of the Google AI Challenge 2010.

For now, the documentation is just a reference. See test/test-uct.lisp for an example.

• [class] uct-node

  A node in the uct tree. Roughly translates to a state in the search space. Note that the state itself is not stored explicity, but it can be recovered by `replaying' the actions from the starting state or by customizing make-uct-node.

• [reader] depth uct-node (:depth = 0)

• [accessor] edges uct-node

  Outgoing edges.

• [accessor] average-reward uct-node (:average-reward = 0)

  Average reward over random playouts started from below this node. See update-uct-statistics and REWARD.

• [class] uct-edge

  An edge in the uct tree. Represents an action taken from a state. The value of an action is the value of its target state which is not quite as generic as it could be; feel free to specialize average-reward for the edges if that's not the case.

• [reader] action uct-edge (:action)

  The action represented by the edge.

• [accessor] from-node uct-edge (:from-node)

  The node this edge starts from.

• [accessor] to-node uct-edge (= nil)

  The node this edge points to if the edge has been visited or nil.

• [function] visited-edges node

• [function] unvisited-edges node

• [generic-function] edge-score node edge exploration-bias

• [generic-function] select-edge node exploration-bias

  Choose an action to take from a state, in other words an edge to follow from node in the tree. The default implementation chooses randomly from the yet unvisited edges or if there is none moves down the edge with the maximum edge-score. If you are thinking of customizing this, for example to make it choose the minimum at odd depths, the you may want to consider specializing REWARD or update-uct-statistics instead.

• [generic-function] outcome->reward node outcome

  Compute the reward for a node in the tree from outcome that is the result of a playout. This is called by the default implementation of update-uct-statistics. This is where one typically negates depending on the parity of depth in two player games.

• [generic-function] update-uct-statistics root path outcome

  Increment the number of visits and update the average reward in nodes and edges of path. By default, edges simply get their visit counter incremented while nodes also get an update to average-reward based on what outcome->reward returns.

• [generic-function] make-uct-node parent edge parent-state

  Create a node representing the state that edge leads to (from parent). Specialize this if you want to keep track of the state, which is not done by default as it can be expensive, especially in light of TAKE-ACTION mutating it. The default implementation simply creates an instance of the class of parent so that one can start from a subclass of uct-node and be sure that that class is going to be used for nodes below it.

• [generic-function] state node parent edge parent-state

  Return the state that corresponds to node. This is not a straightforward accessor unless node is customized to store it. The rest of the parameters are provided so that one can reconstruct the state by taking the action of edge in the parent-state of parent. It's allowed to mutate parent-state and return it. This function must be specialized.

• [generic-function] list-edges node state

  Return a list of edges representing the possible actions from node with state. This function must be customized.

• [generic-function] play-out node state reverse-path

  Play a random game from node with state and return the outcome that's fed into update-uct-statistics. The way the random game is played is referred to as `default policy' and that's what makes or breaks uct search. This function must be customized.

• [function] uct &key root fresh-root-state exploration-bias max-n-playouts

  Starting from the root node, search the tree expanding it one node for each playout. Finally return the mutated root. root may be the root node of any tree, need not be a single node with no edges. fresh-root-state is a function that returns a fresh state corresponding to root. This state will be destroyed unless special care is taken in state.

← ↑ → ↺ λ

3 Metropolis Hastings

[in package MICMAC.METROPOLIS-HASTINGS with nicknames MICMAC.MH]

Generic interface for the Metropolis-Hastings algorithm, also Metropolis Coupled MCMC.

References:

• http://en.wikipedia.org/wiki/Metropolis–Hastings_algorithm

• Markov Chain Monte Carlo and Gibbs Sampling Lecture Notes for EEB 581, version 26 April 2004 c B. Walsh 2004 http://web.mit.edu/~wingated/www/introductions/mcmc-gibbs-intro.pdf

• Geyer, C.J. (1991) Markov chain Monte Carlo maximum likelihood

For now, the documentation is just a reference. See test/test-metropolis-hastings.lisp for an example.

• [class] mc-chain

  A simple markov chain for Metropolis Hastings. With temperature it is suitable for mc3.

• [accessor] temperature mc-chain (:temperature = 1.0d0)

  The PROBABILITY-RATIO of samples is raised to the power of 1 / temperature before calculating the acceptance probability. This effectively flattens the peaks if temperature > 1 which makes it easier for the chain to traverse deep valleys.

• [reader] state mc-chain (:state)

  This is the current sample where the chain is.

• [function] jump-to-sample chain jump &key (result-sample (state chain))

  From the current state of chain make jump (from the current distribution of chain) and return the sample where we landed. Reuse result-sample when possible.

• [generic-function] jump-to-sample* chain jump result-sample

  This function is called by jump-to-sample. It is where jump-to-sample behaviour shall be customized.

• [generic-function] prepare-jump-distribution chain

  Prepare for sampling from the F(X) = Q(SAMPLE->X) distribution. Called by random-jump. The around method ensures that nothing is done unless there was a state change.

• [generic-function] random-jump chain

  Sample a jump from the current distribution of jumps that was computed by prepare-jump-distribution.

• [generic-function] log-probability-ratio chain sample1 sample2

  Return P(sample1)/P(sample2). It's in the log domain to avoid overflows and the ratio part is because that it may allow computational shortcuts as opposed to calculating unnormalized probabilities separately.

• [generic-function] log-probability-ratio-to-jump-target chain jump target

  Return P(target)/P(state) where jump is from the current state of chain to target sample. This can be specialized for speed. The default implementation just falls back on log-probability-ratio.

• [generic-function] log-jump-probability-ratio chain jump target

  Return Q(TARGET->STATE) / Q(STATE->TARGET) where Q is the jump distribution and jump is from the current state of chain to target sample.

• [generic-function] acceptance-probability chain jump candidate

  Calculate the acceptance probability of candidate to which jump leads from the current state of chain.

• [generic-function] accept-jump chain jump candidate

  Called when chain accepts jump to candidate.

• [generic-function] reject-jump chain jump candidate

  Called when chain rejects jump to candidate. It does nothing by default, it's just a convenience for debugging.

• [generic-function] maybe-jump chain jump candidate acceptance-probability

  Randomly accept or reject jump to candidate from the current state of chain with acceptance-probability.

• [generic-function] jump chain

  Take a step on the markov chain. Return a boolean indicating whether the proposed jump was accepted.

• [class] mc3-chain mc-chain

  High probability island separated by low valley make the chain poorly mixing. mc3-chain has a number of hot-chains that have state probabilities similar to that of the main chain but less jagged. Often it suffices to set the temperatures of the hot-chains higher use the very same base probability distribution.

• [generic-function] accept-swap-chain-states mc3 chain1 chain2

  Swap the states of chain1 and chain2 of mc3.

• [generic-function] reject-swap-chain-states mc3 chain1 chain2

  Called when the swap of states of chain1 and chain2 is rejected. It does nothing by default, it's just a convenience for debugging.

• [generic-function] maybe-swap-chain-states mc3 chain1 chain2 acceptance-probability

  Swap of states of chain1 and chain2 of mc3 with acceptance-probability.

• [generic-function] jump-between-chains mc3

  Choose two chains randomly and swap their states with mc3 acceptance probability.

• [class] enumerating-chain mc-chain

  A simple abstract chain subclass that explicitly enumerates the probabilities of the distribution.

• [class] tracing-chain

  Mix this in with your chain to have it print trace of acceptances/rejections.

← ↑ ↺ λ

4 Game Theory

[in package MICMAC.GAME-THEORY]

• [function] find-nash-equilibrium payoff &key (n-iterations 100)

  Find a Nash equilibrium of a zero-sum game represented by payoff matrix (a 2d matrix or a nested list). payoff is from the point of view of the row player: the player who choses column wants to minimize, the row player wants to maximize. The first value returned is a vector of unnormalized probabilities assigned to each action of the row player, the second value is the same for the column player and the third is the expected payoff of the row player in the nash equilibrium represented by the oddment vectors.