create new connection script

src/site/snake/brain-neat.ts

@@ -119,7 +119,7 @@
  *    - use random values connection weights
  *    - all nodes should be marked "enabled"
  *    - all organisms in the initial population will be part of species #1
- *    	- this implies that randomized connection weights should be chosen such that c3*W < δ_t for all organisms
+ *      - this implies that randomized connection weights should be chosen such that c3*W < δ_t for all organisms
  * 3. Training Loop
  *    a) Compute fitness f_i for all organisms[i]
  *    b) Find adjusted f_adj_i based on each organism's species
@@ -129,7 +129,7 @@
  *           - All organisms whose fitness passes the survival threshold (based on f_adj_i) are considered survivors
  *           - Add all survivors to the next generation
  *       - Fertility: fertility[i] = # of offspring from organism i as the mom
- *       	- Organisms from species that have not improved in fitness for stag_lim generations are given fertility[i] = 0 (barred from mating)
+ *          - Organisms from species that have not improved in fitness for stag_lim generations are given fertility[i] = 0 (barred from mating)
  *          - Count the number of fertile survivors
  *          - Compute general, remainder = divmod((pop - # survivors), # fertile survivors)
  *          - Select remainder organisms to be given one bonus fertility (at random or by fitness)
@@ -228,8 +228,8 @@
  *   - Effectively pass data from inputs, through hidden nodes, to outputs
  */
 
-import { edgesToNodes, NodeID, Network, Node, topoSort } from './network';
+import { edgesToNodes, NodeID, Network, Node, topoSort, traceParents, RawEdge } from './network';
-import { keyMax, randint, randomNegPos } from './util';
+import { keyMax, randchoice, randint, randomNegPos, setDifference, setMap } from './util';
 
 interface GeneData {
     innovation: number;
@@ -375,15 +375,6 @@ export function chooseSurvivors(population: Population, fitness: Map<Genome, num
 
 // export function mate(a: Genome, b: Genome): Genome {}
 
-function getGenomeNodeIDs(genome: Genome): Set<NodeID> {
-    const nodeIDs = new Set<NodeID>();
-    for (const gene of genome) {
-        nodeIDs.add(gene.src_id);
-        nodeIDs.add(gene.dst_id);
-    }
-    return nodeIDs;
-}
-
 interface MutateConfig {
     mutate_rate: number; // chance to mutate a gene's weight
     assign_rate: number; // chance to assign instead of uniformly perturb
@@ -440,13 +431,53 @@ export function mutate(genome: Genome, config: MutateConfig): Genome {
     }
 
     if (Math.random() < new_connection_rate) {
-        // create a new connection
+        // create a new connection between two *previously unconnected* nodes
-        // between *previously unconnected* nodes
+        // NOTE: there's some performance stuff here that could definitely be improved
-        // TODO: use traceParents
+
-        // TODO: test traceParents
         const nodes = edgesToNodes(newGenome);
-        const sources = Array.from(nodes.values()).filter(n => n.srcs.size === 0);
+
-        const sinks = Array.from(nodes.values()).filter(n => n.dsts.size === 0);
+        // find nodes that can be connected without creating a cycle
+        // a node that is connected to one of its parents creates a cycle
+        const allNodeIDs = new Set(nodes.keys());
+        const parents = traceParents(nodes);
+        const acyclic = new Map<NodeID, Set<NodeID>>();
+        for (const [nodeID, nodeParents] of parents.entries()) {
+            const nodeParentIDs = setMap(nodeParents, n => n.id);
+            const nodeAcyclic = setDifference(allNodeIDs, nodeParentIDs);
+            acyclic.set(nodeID, nodeAcyclic);
+        }
+
+        // flatten
+        const acyclicConns: { src_id: NodeID, dst_id: NodeID }[] = [];
+        for (const [nodeID, nodeAcyclic] of acyclic) {
+            acyclicConns.push(...setMap(nodeAcyclic, dst => ({ src_id: nodeID, dst_id: dst })));
+        }
+
+        // remove options that are already connected
+        const options: { src_id: NodeID, dst_id: NodeID }[] = [];
+        for (const conn of acyclicConns) {
+            if (newGenome.findIndex(c => c.src_id === conn.src_id && c.dst_id === conn.dst_id) === -1) continue;
+            options.push(conn);
+        }
+
+        // choose a random connection
+        if (options.length === 0) {
+            // TODO: remove this warn once this starts working
+            //       this is mostly a sanity check / useful for metrics
+            console.warn("could not find a valid new connection!");
+        } else {
+            const { src_id, dst_id } = randchoice(options);
+            const newGene = {
+                src_id,
+                dst_id,
+                data: {
+                    innovation: g_innovation_number++,
+                    weight: assign_mag * randomNegPos(),
+                    enabled: true,
+                }
+            }
+            newGenome.push(newGene);
+        }
     }
 
     return newGenome;
@@ -476,6 +507,7 @@ export class Organism {
 
     /** given per-node input activations, computes downstream activations in-place.*/
     think(activations: Map<NodeID, number>) {
+        // TODO: do not follow disabled connections
         for (const id of this.order) {
             const node = this.network.get(id)!;
             if (node.srcs.size === 0) {

src/site/snake/game-engine.ts

@@ -7,6 +7,7 @@ export function randint(low: number, high: number) {
     return Math.floor(Math.random() * range) + low;
 }
 
+/** deprecated, use util.ts instead */
 export function randchoice<T>(arr: T[]) {
     return arr[randint(0, arr.length - 1)]!;
 }

src/site/snake/util.ts

@@ -9,6 +9,10 @@ export function randint(low: number, high: number) {
     return Math.floor(Math.random() * range) + low;
 }
 
+export function randchoice<T>(arr: T[]) {
+    return arr[randint(0, arr.length - 1)]!;
+}
+
 export function keyMax<T>(values: Iterable<T>, keyFunc: (v: T) => number) {
     let best = null;
     for (const value of values) {