speedyjeff/prey

Predator v Prey

Predator (blue) versus prey (yellow) simulator. This is inspired by AI Battles.

The AI are driven by a neural networks that choose the direction and speed of movement. In the initial configuration the predators (blue) hunt and attack the prey (yellow). A series of experiments were run to guage the fitness of these networks (compared with a random movement AI). The results show that the average lifetime of a predator is 40-70% longer than moving randomly.

Basic Rules

The basic rules for the simulatio are that predators must eat prey to survive, and prey try to survive as long as they can.

Predators

• Movement consume energy, standing still consumes energy
• Must consume prey in order to gain energy and reproduce
• Once energy is depleted, they will die
• Touching a prey is lethal
• Eating when not fully digested their previous meal will gain energy, but not contribute to their next reproduction
• The field of vision is narrow and long

Prey

• Movement consumes energy, standing still replenishes energy
• Reproductio happens as long as they stay alive long enough
• Once energy is depleted, they cannot move
• The field of vision is wide and short

Optional rules include allowing prey to fight back and kill predators, requiring N hits to deal a lethal hit, etc.

Mechanics

The AI are governed by a set of meters that are adjusted for every update. An update is designed to happen 10 times a second. Predators and prey have specific costs/update for each of these meters.

• Energy - needed to move/survive
• Digestion - (predator) when still digesting, eating more prey does not lower the reproduction meter
• Reproduction
  • (predator) adjusts when digesting prey
  • (prey) adjusted per update
• Attack - a cool down between each attack

Each AI also has a Health meter, which is only adjusted if attacked by an AI of the opposite type.

The field of view is generated by extending rays out in front of the AI and returning the objects that are the closest.

Movement and speed are determined by two neural networks. These neural networks are trained by actions that the AI take. At first, the actions are largely random, but as the AI survives the actions are reinforced by actions that yield longer life. New AI have a chance to inhert traits from their parents or the best surviving AI.

Experiments

Experiments are run to track if neural network trained AI can have extended average lifetimes to random AI. The termination conditions for the experiments are either of the populations are extinct, or a maximum lifetime is reached (maximum number of updates).

The minimum / maximum lifetimes: (without being killed)

• Predator - 200 thru 1000 (even further if consuming prey)
• Prey - maxium lifetime (per the basic rules, prey will stay still once energy is depleted)

Small experiment

The small experiments have a limit of ~1,000 AI. This is a great experiment to visualize how the modes evolve and interact.

Configuration:

• Width - 6,400
• Height - 6,400
• Predator (initial/max) - 50/170
• Prey (initial/max) - 300/800
• Combat - variable (peaceful or aggressive)
• Random - variable
• Iterations - 10
• Max Lifetime - 5,000

./simulationcli -width 6400 -height 6400 -initialpred 50 -initialprey 300 -maxpred 170 -maxprey 800 -combat $0 -random $1 -it 10 -maxl 5000

Predator (peaceful) v Prey (peaceful)

This configuration is useful for a baseline, it will help determine local minimums for the predators as they require combat (eating) to survive. The average lifetime of predator (neural) was 70% higher than predator (random). The average lifetime for predators (neural) were 586-589 updates, where as predator (random) was 333-334 updates.

data

Predator (neural) v Prey (neural)

Iteration	Lifetime	AlivePredators	AlivePrey	BestPredatorLifetime	BestPredatorGeneration	BestPreyLifetime	BestPreyGenertation	AvgPredatorLifetime	AvgPreyLifetime
0	673	0	800	672	0	1	0	421	0
1	698	0	800	696	1	0	0	606	0
2	675	0	800	672	2	0	0	589	0
3	689	0	800	688	3	0	0	616	0
4	680	0	800	680	4	0	0	560	0
5	710	0	800	708	5	0	0	618	0
6	671	0	800	669	6	0	0	626	0
7	712	0	800	710	7	0	0	609	0
8	676	0	800	676	8	1	0	611	0
9	703	0	800	702	9	0	0	606	0

Predator (random) v Prey (neural)

Iteration	Lifetime	AlivePredators	AlivePrey	BestPredatorLifetime	BestPredatorGeneration	BestPreyLifetime	BestPreyGenertation	AvgPredatorLifetime	AvgPreyLifetime
0	359	0	800	359	0	0	0	334	0
1	362	0	800	360	1	0	0	338	0
2	355	0	800	352	2	1	0	334	0
3	365	0	800	364	3	1	0	335	0
4	364	0	800	364	4	1	1	334	0
5	354	0	800	353	5	0	0	331	0
6	370	0	800	370	0	0	0	335	0
7	355	0	800	355	1	1	0	332	0
8	360	0	800	359	2	0	0	335	0
9	365	0	800	363	3	0	0	335	0

Predator (neural) v Prey (random)

Iteration	Lifetime	AlivePredators	AlivePrey	BestPredatorLifetime	BestPredatorGeneration	BestPreyLifetime	BestPreyGenertation	AvgPredatorLifetime	AvgPreyLifetime
0	676	0	800	668	0	0	0	429	0
1	674	0	800	674	1	0	0	614	0
2	701	0	800	690	2	0	0	620	0
3	699	0	800	692	3	0	0	613	0
4	700	0	800	684	4	0	0	611	0
5	700	0	800	694	5	0	0	609	0
6	689	0	800	672	6	0	0	604	0
7	687	0	800	686	7	0	0	601	0
8	691	0	800	688	8	0	0	588	0
9	706	0	800	696	9	1	0	604

Predator (random) v Prey (random)

Iteration	Lifetime	AlivePredators	AlivePrey	BestPredatorLifetime	BestPredatorGeneration	BestPreyLifetime	BestPreyGenertation	AvgPredatorLifetime	AvgPreyLifetime
0	359	0	801	352	0	0	0	330	0
1	387	0	800	372	1	1	0	334	0
2	367	0	800	359	2	0	0	335	0
3	361	0	800	352	3	0	0	334	0
4	368	0	800	354	4	0	0	336	0
5	371	0	800	363	5	1	0	333	0
6	363	0	800	357	6	0	0	333	0
7	381	0	800	372	0	0	0	335	0
8	353	0	800	351	1	0	0	332	0
9	369	0	800	356	2	0	0	335	0

Predator (aggressive) v Prey (peaceful)

This is the default configuration, the predators hunt the prey. The predator lifetime saw a 40% increase when using the predator (neural) over predator (random). The average lifetime of predator (neural) was 670-818 updates, where as predator (random) was 514-559 updates.

data

Predator (neural) v Prey (neural)

Iteration	Lifetime	AlivePredators	AlivePrey	BestPredatorLifetime	BestPredatorGeneration	BestPreyLifetime	BestPreyGenertation	AvgPredatorLifetime	AvgPreyLifetime
0	5000	170	800	2490	5	3288	2	396	238
1	5000	170	800	1912	11	4495	3	651	560
2	4086	0	800	1674	12	3363	5	841	610
3	4221	0	800	1689	14	3335	7	855	519
4	4597	0	800	1681	17	3956	8	842	826
5	5000	170	799	1788	18	3899	9	469	302
6	5000	166	800	2004	25	4306	11	486	383
7	3905	0	800	1414	26	3705	13	774	431
8	4235	0	800	1614	27	3211	14	860	585
9	5000	166	801	1551	8	4068	15	535	339

Predator (random) v Prey (neural)

Iteration	Lifetime	AlivePredators	AlivePrey	BestPredatorLifetime	BestPredatorGeneration	BestPreyLifetime	BestPreyGenertation	AvgPredatorLifetime	AvgPreyLifetime
0	5000	170	797	2020	5	2542	9	545	251
1	5001	157	800	2302	10	2648	12	518	243
2	5001	169	790	2038	14	3212	6	512	247
3	5000	166	799	2305	23	3939	9	511	270
4	5000	169	800	2013	32	3409	11	512	259
5	5000	170	794	1608	39	3469	19	516	269
6	5000	170	797	1638	47	3365	23	494	283
7	5000	169	800	1904	54	3145	25	545	260
8	5001	168	799	1786	64	2932	26	492	285
9	5000	169	799	1762	72	3610	27	499	273

Predator (neural) v Prey (random)

Iteration	Lifetime	AlivePredators	AlivePrey	BestPredatorLifetime	BestPredatorGeneration	BestPreyLifetime	BestPreyGenertation	AvgPredatorLifetime	AvgPreyLifetime
0	5002	169	800	2278	1	4386	1	675	414
1	5000	170	800	1688	4	3116	7	630	321
2	5000	170	800	2314	8	3541	11	726	353
3	5003	170	800	2645	9	4481	1	943	542
4	5004	170	801	2337	13	4445	0	769	436
5	5001	169	799	2788	18	4267	4	919	519
6	5002	170	800	3136	19	4011	6	904	528
7	5005	167	800	3101	21	4517	10	922	522
8	5008	168	800	2834	23	4394	15	958	526
9	5002	168	799	2625	26	4093	18	740	336

Predator (random) v Prey (random)

Iteration	Lifetime	AlivePredators	AlivePrey	BestPredatorLifetime	BestPredatorGeneration	BestPreyLifetime	BestPreyGenertation	AvgPredatorLifetime	AvgPreyLifetime
0	5007	169	800	2371	3	2679	6	546	270
1	5005	170	797	2218	10	3076	9	534	273
2	5001	169	800	2279	11	3016	5	545	269
3	5001	170	800	2168	18	2712	7	562	263
4	5002	169	797	1927	23	3127	13	577	259
5	5000	170	800	1765	26	3090	20	557	272
6	5004	170	801	1996	27	2644	24	579	265
7	5002	169	797	2214	28	2904	0	549	274
8	5006	167	798	2096	31	2574	5	566	264
9	5000	170	799	2074	7	3229	8	583	256

Predator (peaceful) v Prey (aggressive)

This configuration has a double negative for predators, they are not able to reproduce (as they cannot eat) and prey are agressive. There was a significant drop (2x) in average lifetimes for predators compared with the baseline.

data

Predator (neural) v Prey (neural)

Iteration	Lifetime	AlivePredators	AlivePrey	BestPredatorLifetime	BestPredatorGeneration	BestPreyLifetime	BestPreyGenertation	AvgPredatorLifetime	AvgPreyLifetime
0	486	0	800	486	0	1	0	211	0
1	461	0	800	461	1	0	0	138	0
2	649	0	800	648	2	0	0	209	0
3	676	0	800	676	3	0	0	235	0
4	512	0	800	512	4	1	0	192	0
5	698	0	800	698	5	0	0	317	0
6	642	0	800	642	6	0	0	199	0
7	652	0	800	652	7	0	0	182	0
8	439	0	800	438	8	0	0	146	0
9	666	0	800	666	9	0	0	193	0

Predator (random) v Prey (neural)

Iteration	Lifetime	AlivePredators	AlivePrey	BestPredatorLifetime	BestPredatorGeneration	BestPreyLifetime	BestPreyGenertation	AvgPredatorLifetime	AvgPreyLifetime
0	347	0	800	346	0	0	0	171	0
1	351	0	800	350	1	1	0	184	0
2	358	0	800	358	2	0	0	231	0
3	342	0	801	342	3	0	0	182	0
4	362	0	800	361	4	0	0	186	0
5	347	0	800	347	5	0	0	180	0
6	347	0	800	346	6	0	0	177	0
7	342	0	800	342	7	0	0	155	0
8	340	0	800	340	8	0	0	189	0
9	336	0	800	335	9	0	0	166	0

Predator (neural) v Prey (random)

Iteration	Lifetime	AlivePredators	AlivePrey	BestPredatorLifetime	BestPredatorGeneration	BestPreyLifetime	BestPreyGenertation	AvgPredatorLifetime	AvgPreyLifetime
0	380	0	800	376	0	0	0	161	0
1	310	0	801	308	1	1	2	143	0
2	612	0	800	604	2	0	0	177	0
3	413	0	800	406	3	1	2	144	0
4	521	0	800	519	4	1	3	174	0
5	395	0	800	394	5	0	0	170	0
6	692	0	800	691	6	2	2	172	0
7	607	0	800	602	0	1	3	165	0
8	523	0	800	514	1	0	0	182	0
9	540	0	800	539	2	0	0	151	0

Predator (random) v Prey (random)

Iteration	Lifetime	AlivePredators	AlivePrey	BestPredatorLifetime	BestPredatorGeneration	BestPreyLifetime	BestPreyGenertation	AvgPredatorLifetime	AvgPreyLifetime
0	341	0	800	337	0	1	2	150	0
1	349	0	800	344	1	0	0	183	0
2	353	0	800	348	2	0	0	181	0
3	327	0	800	322	3	0	0	172	0
4	352	0	800	345	4	0	0	162	0
5	322	0	800	320	5	0	0	144	0
6	349	0	800	339	6	1	0	176	0
7	354	0	800	349	7	0	0	185	0
8	336	0	800	331	8	0	0	165	0
9	335	0	800	331	9	0	0	177	0

Predator (aggresive) v Prey (aggressive)

This is the purge mode, everyone is attacking everyone. The prey seem to have the advantage as they have the numbers and significantly reduce (5x) the length of the game, espeically once the prey (neural) model is trained.

data

Predator (neural) v Prey (neural)

Iteration	Lifetime	AlivePredators	AlivePrey	BestPredatorLifetime	BestPredatorGeneration	BestPreyLifetime	BestPreyGenertation	AvgPredatorLifetime	AvgPreyLifetime
0	4543	0	800	809	2	4199	0	298	1618
1	5001	10	800	1055	16	4037	1	315	445
2	3621	0	800	1289	17	2737	1	336	416
3	913	0	800	486	18	865	2	172	169
4	649	0	800	648	0	331	3	193	105
5	717	0	800	716	1	517	4	260	102
6	726	0	800	726	2	408	5	240	88
7	729	0	800	729	3	345	7	254	81
8	770	0	800	770	0	326	8	248	98
9	702	0	800	702	1	470	10	227	104

Predator (random) v Prey (neural)

Iteration	Lifetime	AlivePredators	AlivePrey	BestPredatorLifetime	BestPredatorGeneration	BestPreyLifetime	BestPreyGenertation	AvgPredatorLifetime	AvgPreyLifetime
0	4448	0	800	560	11	4127	0	284	1102
1	5000	48	800	912	32	3960	3	344	615
2	5000	14	800	794	38	3801	6	324	562
3	5000	17	800	822	56	4656	8	348	816
4	5000	100	800	800	60	4757	1	322	665
5	847	0	800	510	61	694	3	204	202
6	1772	0	800	422	65	1428	4	197	202
7	519	0	800	436	66	482	0	198	118
8	681	0	800	427	67	594	1	209	90
9	829	0	800	416	0	722	3	176	232

Predator (neural) v Prey (random)

Iteration	Lifetime	AlivePredators	AlivePrey	BestPredatorLifetime	BestPredatorGeneration	BestPreyLifetime	BestPreyGenertation	AvgPredatorLifetime	AvgPreyLifetime
0	5000	29	800	549	4	4977	0	203	1577
1	720	0	800	480	5	669	1	193	188
2	713	0	801	475	6	596	2	205	183
3	847	0	801	430	7	653	4	200	149
4	730	0	800	438	8	572	6	231	162
5	661	0	800	448	9	470	7	221	137
6	538	0	800	468	0	435	8	195	135
7	780	0	800	465	2	704	0	235	183
8	1837	0	802	436	3	1816	1	211	495
9	866	0	800	406	4	591	3	233	181

Predator (random) v Prey (random)

Iteration	Lifetime	AlivePredators	AlivePrey	BestPredatorLifetime	BestPredatorGeneration	BestPreyLifetime	BestPreyGenertation	AvgPredatorLifetime	AvgPreyLifetime
0	1019	0	800	457	1	869	1	216	155
1	2039	0	800	508	6	1895	2	222	395
2	875	0	800	472	7	697	4	231	184
3	1221	0	800	420	8	1114	5	214	252
4	1680	0	800	526	12	1463	0	233	219
5	930	0	800	401	0	870	0	221	166
6	635	0	800	432	1	598	1	228	147
7	1778	0	800	467	2	1558	3	220	363
8	526	0	800	471	4	394	5	224	136
9	1238	0	800	508	0	1046	6	245	233

Large experiment

The large experiments have a limit of ~16,000 AI, an increase of 16x over the small experiment.

Configuration:

• Width - 16,000
• Height - 16,000
• Predator (initial/max) - 800/800
• Prey (initial/max) - 4800/13000
• Combat - variable (peaceful or aggressive)
• Random - variable
• Iterations - 5
• Max Lifetime - 10,000

./simulationcli -width 16000 -height 16000 -initialpred 800 -initialprey 4800 -maxpred 800 -maxprey 13000 -combat $0 -random $1 -it 10 -maxl 10000 

Predator (aggressive) v Prey (peaceful)

This larger configuration provides more space, more predators, and more prey. Predator lifetime saw a 75% increase when using the predator (neural) over predator (random). The average lifetime of predator (neural) was 1183 updates, where as predator (random) was 674 updates. In addition, prey (neural) lifetime was 55% longer than prey (random).

data

Predator (neural) v Prey (neural)

Iteration	Lifetime	AlivePredators	AlivePrey	BestPredatorLifetime	BestPredatorGeneration	BestPreyLifetime	BestPreyGenertation	AvgPredatorLifetime	AvgPreyLifetime
0	10000	800	11922	8127	1	5746	0	854	323
1	10000	800	12997	9898	2	9345	2	1187	531
2	10000	800	12997	8956	3	9664	3	1188	479
3	10000	800	12998	8881	4	9939	4	1283	644
4	10000	800	13000	9705	5	9810	6	1405	612

Predator (random) v Prey (random)

Iteration	Lifetime	AlivePredators	AlivePrey	BestPredatorLifetime	BestPredatorGeneration	BestPreyLifetime	BestPreyGenertation	AvgPredatorLifetime	AvgPreyLifetime
0	10000	800	12999	6414	5	5428	11	719	327
1	10000	800	13000	3948	18	4688	21	667	341
2	10000	800	12997	5638	29	5740	32	656	335
3	10000	800	13000	3966	42	5280	37	652	328
4	10000	798	12991	4204	55	5053	45	680	333

Predator (aggresive) v Prey (aggressive)

This configuration continues to be a struggle for predators, as prey are aggressive and attack back. Predator (neural) and predator (random) were within 8%, with random trending better. Prey (neural) average lifetime was nearly 4x higher than prey (random).

data

Predator (neural) v Prey (neural)

Iteration	Lifetime	AlivePredators	AlivePrey	BestPredatorLifetime	BestPredatorGeneration	BestPreyLifetime	BestPreyGenertation	AvgPredatorLifetime	AvgPreyLifetime
0	8609	0	13000	712	0	7977	1	199	1327
1	428	0	13000	428	1	333	2	103	77
2	405	0	13000	388	2	372	0	100	77
3	10000	104	13000	786	3	9994	1	208	2408
4	733	0	13000	706	4	720	0	135	83

Predator (random) v Prey (random)

Iteration	Lifetime	AlivePredators	AlivePrey	BestPredatorLifetime	BestPredatorGeneration	BestPreyLifetime	BestPreyGenertation	AvgPredatorLifetime	AvgPreyLifetime
0	1084	0	13000	522	0	954	0	162	145
1	1335	0	13000	554	1	1320	1	161	172
2	4201	0	13000	470	3	4161	2	158	391
3	996	0	12999	592	4	945	3	162	153
4	869	0	13000	575	5	789	4	162	139

Long running

This experiment was to see if over the long term who wins: predators or prey. When predator (neural) and prey (nerual) are active, it was a toss up with prey winning more often. When predator (random) and prey (random) are active, the predators win. If you visualize the difference between predator (neural) and predator (random), predator (random) holds even distribution across the board (eg. playing zone defense). Whereas predator (neural) sweeps the prey away and follows the respan around the board. The later approach can cause mass extinction if there are no predators in the zone where prey are spawning.

Configuration:

• Width - 6,400
• Height - 6,400
• Predator (initial/max) - 50/170
• Prey (initial/max) - 300/800
• Combat - Predator
• Random - variable
• Iterations - 5
• Max Lifetime - 100,000

simulationcli.exe -width 6400 -height 6400 -initialpred 50 -initialprey 300 -maxpred 170 -maxprey 800 -combat 1 -random 0 -it 5 -maxl 100000 

Predator (aggressive) v Prey (peaceful)

data

Predator (neural) v Prey (neural)

Iteration	Lifetime	AlivePredators	AlivePrey	BestPredatorLifetime	BestPredatorGeneration	BestPreyLifetime	BestPreyGenertation	AvgPredatorLifetime	AvgPreyLifetime
0	52413	167	0	2084	9	5415	41	466	321
1	40462	0	800	3771	53	16975	45	782	1062
2	3214	0	800	1484	55	2572	46	790	616
3	26737	156	0	2528	68	8711	47	486	461
4	5480	0	800	1609	73	4887	48	841	1169

Predator (random) v Prey (random)

Iteration	Lifetime	AlivePredators	AlivePrey	BestPredatorLifetime	BestPredatorGeneration	BestPreyLifetime	BestPreyGenertation	AvgPredatorLifetime	AvgPreyLifetime
0	52271	168	0	2134	158	4265	57	475	315
1	43907	163	0	2287	283	4086	69	480	317
2	26823	165	0	2236	304	3870	84	470	311
3	31292	169	0	1787	323	4139	148	473	322
4	21914	166	0	1961	364	4727	161	483	308

Next steps

There is a lot more tuning possible in this experiment.
A few areas of focus:

• nerf prey attacks (as they currently are like a hord of unstoppable zobmies and overwhelm the predators)
• further tuning of AI meters
• allow for longer sight by predators
• perform a tuning experiment to find the cross over points of where predators dominate, prey dominate, and a balance

Trying

This project is build able with Visual Studio 2022 or later. Below are the steps to reproduce these experiments.

Visual representation (Windows only)

simulation.exe

Cross platform and on the command line.

./simulationcli.exe

This is all the available options for trying out the simulation from the command line.

./simulationcli.exe -help
./simulationcli
  -(help)             : this help
  -(w)idth            : the 2d width of the environment (default: 6400)
  -(hei)ght           : the 2d height of the enviorment (default: 6400)
  -(initialPred)ators : the initial number of predators at start (default: 50)
  -(maxPred)ators     : the max number of predators (default: 170)
  -(initialPrey)      : the initial number of prey at start (default: 300)
  -(maxPrey)          : the max numbre of prey (default: 800)
  -(c)ombat           : allowed combat (0: none, 1:Predators, 2:Prey, 3:Predators+Prey) (default: 1)
  -(r)andom           : random movement (0: none, 1:Predators, 2:Prey, 3:Predators+Prey) (default: 0)
  -(it)erations       : number of training runs (a run stops when either Predators or Prey are extinct) (default: 10)
  -(maxl)ifetime      : maximum number of updates per iteration (default: 5000

Initialization

git submodule init
git submodule update